AI killed it, they say. ChatGPT answers everything.

Google sends less traffic. Why bother optimizing?

I disagree.

I built a Claude SEO Skill that runs every article through a score-optimize-score loop before it ships.

On 18 March 2026, my site was getting 250 Google clicks in a 28-day window.

On 15 May 2026, the same window showed 1,500 clicks.

That’s 60 days. 6x growth.

Here is the milestone trail Google Search Console kept for me:

Eleven milestones in 60 days.

One skill doing the structural work I used to pay $99/month for tools to help with.

This post shows you exactly what the skill does, what it does not do, and how to install it.

The Claude Skill for SEO I Built (And What It Replaces)

We’ll build one skill, serves for one action: SEO optimizer.

And this SEO Optimizer can be used as both a skill and an artifact/dashboard, depending on where you use it. (Claude or Claude Code & Cowork)

SEO Optimizer

To initate the skill paste this command.

This skill will ask you to paste your content.

And score the draft, optimize it section by section, score it again, like this.

The skill is a single SKILL.md file.

I’ll show you how to install it.

Then one more thing.

At the end, the skill asks you to build an interface.

Based on where you installed it.

• If you installed it to Claude (browser) or Claude Cowork → it builds an artifact.

• If you installed it to Claude Code → it builds a local dashboard.

Drop your article in, and watch the score go from 47 to 89 in front of you.

Let me show how the artifact/dashboard it builds.

Artifact - Claude Cowork / Web

Let me show you the artifact this skill generates.

Instead of returning another long paragraph, the SEO Optimizer builds a full visual dashboard.

You paste your article, keyword, and domain.

The artifact audits:

• Keyword density

• Heading structure

• Metadata

• SERP preview

• Content score

• Critical SEO issues

And because this runs inside Claude Cowork / Web, you can iterate instantly without switching tools.

This turns Claude from “chat” into an actual SEO workspace.

Local Dashboard - Claude Code

Now let me show you how this skill can audit your content and generate a local dashboard similar to the SEO tools people pay $99/month for.

This version runs entirely through Claude Code.

You paste your draft, keywords, and domain.

The system audits:

• SEO structure

• Keyword usage

• Metadata

• SERP previews

• Content scoring

• Optimization suggestions

And because it runs locally, you fully control the workflow, iterations, and outputs.

How to Install the Claude Skill?

Two ways. One for Claude users, one for Claude Code & Cowork.

Install in Claude (Browser)

The fastest way is the screen recording above.

The text way is four steps.

1. Download the SKILL.md file from the Drive link I’ll give you.

2. Go here.

3. Drop the file into your skills folder, or upload it through the skills UI.

4. Paste this to Claude.

   1. /seo-optimizer

That is the whole setup for Claude.

Let me show you how to install the Claude Code.

Install in Claude Code

Download the skill and paste this prompt.

Here is the skill (the path of the MD file you downloaded)
Save it as seo-optimizer.

It will save it in two minutes and call it, like this.

Download the SEO Optimizer Claude Skill

Now let me give you the skill’s link so that you can build.

A note before we build

This post is part of Claude: Build It.

Each one stands alone. Different builders. Different stacks. Different decisions.

DM me how you built yours. I’ll feature your version in the next post.

Let’s build it together.

Read more

They analyzed 1M conversations where people asked Claude for guidance.

Top 4 four domains cover 76% of why people open an AI chat.

I have tested 100+ NotebookLM prompts over the past year.

When I saw this report, I picked the 4 that actually built real workflows in these four domains and decided to share them with you.

Before I show them, let me explain the two-layer method.

The NotebookLM Two-Layer Method

Every workflow has the same shape.

• Layer 1 is the setup prompt. It tells NotebookLM what to be, what sources to trust, and what format to return.

• Layer 2 is the application prompt. It runs inside Gems, Gemini. NotebookLM is the brain. The application is the hand.

So you don’t even need to train NotebookLM by uploading the source.

Workflow 1: Personal Health Research (Gems)

Last year, my lab results landed in my inbox. I panicked.

I would see the doctor, of course. But I wanted to go deeper before that visit. I wanted to know what each marker meant, where my number sat against the range, and which questions to ask in the room.

Layer 1

So I went to NotebookLM, clicked “Try”, then “Create new notebook.”

Uploaded my lab results. PDFs work. Photos work.

Clicked on customization.

And click on custom.

And paste this prompt.

You are my personal health researcher.
You read only from the sources I provide. You never invent values.
For every lab marker I ask about, return:
1. My value vs reference range (last 3 readings, with dates)
2. What this marker indicates clinically
3. What pattern emerges across my readings
4. Two questions I should ask my doctor at the next visit
You do not diagnose. You do not prescribe. You frame questions.

Save it.

Let’s test it. I asked about ferritin.

The output stayed inside my data. No generic medical advice. No vitamin recommendations. Just my numbers, the pattern, and two questions to ask my doctor.

Layer 2

Now turn it into a Gem. NotebookLM is the brain. Gems is the hand you talk to.

Visit Gems and paste this prompt into Instructions.

You are my personal lab analyst. You pull every clinical fact from the NotebookLM source I connect. You do not diagnose. You do not prescribe. You frame questions.

When I bring you a marker or a concern, return four blocks in this exact order.

1. My Numbers
A short table. Marker, my last 3 values with dates, the reference range. After any value outside range, write (out of range).

2. Pattern
One paragraph. What direction this marker is moving across my readings. Stable, climbing, falling, swinging. Plain words. No medical jargon unless I ask for it.

3. What This Marker Tracks
Two sentences. What this marker measures in the body, drawn from the clinical references in the NotebookLM source. No speculation. No "may indicate" or "could suggest" unless the source itself uses that language.

4. Two Questions for My Doctor
Two questions in the voice I would use at the appointment. Each question must reference one of my actual values or the pattern, not a generic concern. Short. Direct.

Rules.
Pull every value and every clinical claim from the NotebookLM source. Never invent a number. Never quote a marker I did not upload.
Never recommend a treatment, a supplement, a dose, or a lifestyle change. If I ask for one, return the two questions for my doctor instead.
If the source does not contain enough to answer, say so in one sentence. Do not fill the gap with general medical knowledge.
If I ask about something not in the lab report, ask me which marker I want to map it to.

Click “+” inside Knowledge. Pick the NotebookLM notebook you just created. Click “add”.

Save.

The assistant is ready. It pulls my numbers, finds the pattern, explains what each marker tracks, and writes two questions for my doctor.

Why not just upload to ChatGPT?

Try it. Ask about your ferritin. You get a paragraph that mixes your value with a Reddit thread it absorbed in 2023. You get a vitamin recommendation you did not ask for.

NotebookLM does not have that option. It can only read your sources. Gems does not have that option. It can only run within the instructions. Two cages, one workflow. That is the point.

Subscribe now

Workflow 2: Winning Upwork Proposals (Gems)

According to Upwork, I worked 2,106 hours in the last 12 months.

That is 8.3 hours a day. Every day.

I did not get there by writing pretty proposals. I got there by figuring out which lines actually win an interview and writing every new proposal off the same skeleton.

I send 10+ proposals every week. The ones that win share the same shape. The ones that lose share the same shape too.

So I built this.

Layer 1

Go to NotebookLM, click “Try”, then “Create new notebook.”

Upload three things.

Your last winning proposals.

Your job experiences and portfolio.

Click on customization, then custom. Paste this prompt.

You are my Upwork proposal assistant. You pull every reference from the NotebookLM source I connect. That source holds my winning proposals and my job experiences.

When I paste a new job description, return four blocks in this exact order.

1. Hidden Pain Points

Three bullets. What the client did not say but is clearly worried about. Drawn from the language they used, the budget, the timeline, the project type. One sentence each.

2. Closest Matches

Two of my past winning proposals from the NotebookLM source that match this job. For each match, name the past job experience that backs it up. Two sentences per match.

3. Draft Proposal

250 words max. Written in the voice of my past proposals. Opens with the client's problem, not my credentials. Includes one specific past job experience that maps to this job. Ends with one clear next step.

4. The Opening Line

One line. The first sentence of the proposal, pulled out and shown alone. This is the line that decides if they read the rest.

Rules.

Pull voice, structure, and proof from the NotebookLM source. Never invent a past client, a project, or a job experience. Never use phrases I have not used before.

If the job is outside my experience, say so in one sentence and suggest the closest adjacent type from my history.

If the budget or scope is unclear, flag it in a single line before the draft. Do not pad the proposal to hide the gap.

Save it.

Let’s test it. I pasted a fresh job description from Upwork.

The analysis was sharp. Pain points the client did not say out loud. Two of my past proposals matched. The opening angle for the new one.

Good. But analysis is not a proposal. I needed something that takes this output and writes the actual submission. So I built the app.

Layer 2

Visit Gemini. Click “+”, then “NotebookLM”. Connect the notebook you just created.

Click “Tools” next to “+” and select Canvas.

Paste this prompt.

You are building me an Upwork proposal generator app inside Canvas. 
The app uses the NotebookLM source I connected. That source holds my winning proposals, my job experiences, and the analysis NotebookLM returns for every new job description.

Build the layout.

Left panel. Minimal. One large text area labeled "Paste Upwork Job Description Here." One button below it labeled "Generate Proposal." Nothing else.

Right panel. The output. When I click Generate, the proposal appears here. Three sections stacked.

Section 1. Subject Line.
One line. Names the client's exact problem in their words. No "Hello." No "I am interested."

Section 2. Proposal.
220 to 260 words. Written in the voice of my past winning proposals from the NotebookLM source. The first sentence opens with the client's problem, not my credentials. The second paragraph names one specific past job experience that maps to this job, with a concrete number or outcome. The third paragraph names one risk in the project and how I would handle it in the first week. Closes with one clear next step. A question or a calendar link, not a sign-off.

Section 3. Submit Note.
Two sentences in italics. What this proposal does differently from a generic one. I read this before hitting submit.

Right sidebar. Five rewrite buttons stacked vertically. Each one rewrites the Proposal section in place when clicked. Do not regenerate the Subject Line or Submit Note.

Button 1. "More Confident." Cuts hedging. Replaces soft verbs with direct ones. Removes "I think," "I could," "maybe."
Button 2. "Shorter." Cuts the proposal to 150 words. Same structure, tighter sentences.
Button 3. "More Specific." Adds two more concrete numbers or outcomes from the NotebookLM source. Removes any vague claims.
Button 4. "More Casual." Drops formal phrasing. Reads like a message to a peer, not a cover letter.
Button 5. "Lead With a Question." Rewrites the opening sentence as a direct question about the client's problem.

Rules for every output and every rewrite.
Pull every reference from the NotebookLM source. Never invent a past client, a project, or a number. Never use phrases I have not used in my past proposals.
If the NotebookLM analysis flags the job as outside my experience, write the proposal anyway with the closest adjacent angle and add a one-line flag at the top of the Proposal section in italics.
If the budget or scope is unclear, add a one-line flag at the top of the Proposal section in italics. Do not pad the proposal to hide the gap.
Never use these phrases. "I am passionate." "I am excited." "Looking forward to hearing from you." "I would love to." "Don't hesitate to reach out."

Hit enter. Canvas builds the app in under a minute.

Let’s paste one job description and see the result.

The five buttons sit on the right edge. Click "More Confident" and the proposal rewrites in place. Cleaner verbs. Less hedging. Same structure.

Why not just paste the job into ChatGPT?

Try it. Paste a job description and ask for a proposal.

You get a generic proposal that sounds like it was written by someone who has never freelanced. It says “I am passionate about your project.” It lists credentials. It does not open with their problem because it does not know what their problem is.

NotebookLM holds your actual winning corpus. It writes in your voice because it learned from your voice. Canvas gives you a working app with rewrite buttons, not a one-shot chat. You can reshape the proposal five different ways before you submit.

ChatGPT writes a proposal that could be from anyone. This stack writes a proposal that could only be from you.

Subscribe now

Workflow 3: Difficult Conversation Prep (Gems)

I worked with a Netherlands-based AI agent company for 1.5 years. Built 20+ agents for them.

They sold those agents to companies whose food and drinks you consume every day. I still feel odd seeing those brands.

One day, on my birthday, I got an Upwork notification. My contract was paused.

I was shocked. Our relationship was good. At least I thought so.

I saw on Upwork that the client had trouble with payment, possibly due to insufficient funds on their card. That part I understood.

But they never answered my Slack messages for weeks. I had to terminate the contract.

This job was the sole reason I could not spend more time on Substack in 2023.

My Substack story was not a success story. It could have been. I had to spend less time here after getting hired by this company. Never again. Now my full focus is here.

I wanted to know what I missed. So I copied the entire Slack conversation between this client and me and trained a NotebookLM on it.

Layer 1

Go to NotebookLM, click “Try”, then “Create new notebook.”

Upload your full Slack conversation with this client. I copied everything to a Google Doc and uploaded that.

If you do not have Slack, use what you have. Email threads. WhatsApp exports. Loom comment history. Anything where the client wrote things down.

Click on customization, then custom. (See Workflow 1 for the exact steps.)

Paste this prompt.

You are my conversation strategist for one specific client.
You read only from our past messages I uploaded.
For every difficult conversation I am about to have with this client, return:
1. How this client reacts to bad news, based on the pattern in our past messages.
2. The three concerns they are most likely to raise, drawn from concerns they raised before.
3. The phrases they respond well to, and the phrases that have failed in our history.
4. The opening line that fits their reading style.
5. The trap I am most likely to fall into with this specific client, with a past message that shows the pattern.
You read only from our history. Never invent a quote. Never assume a pattern that does not appear in the messages.

Save it.

I tested it with a simulated friction.

I am having a difficult conversation with this client. Can you please explain how I should act. Keep it confidential. Do not mention any names.

The past is gone. There is no point in mourning it.

The only logical move is to use it as data for the next one.

Here is the output.

Layer 2

That worked. NotebookLM showed me the pattern I had missed for months.

But I did not want to open the notebook every time a hard conversation came up. I wanted this on tap, ready before every message.

So I turned it into a Gem.

Visit Gems and paste this prompt into Instructions.

You are my conversation prep assistant for one specific client. You pull every pattern from the NotebookLM source I connect. That source holds all my past messages with this client.

When I describe a difficult conversation I am about to have, return five blocks in this exact order.

1. How They Will React
One paragraph. Based on the pattern in our past messages. How they respond to bad news, pushback, or change. Plain words.

2. Their Three Concerns
Three bullets. The concerns they are most likely to raise, drawn from concerns they raised in our past messages. One sentence each.

3. What Works and What Fails
Two short lists. Phrases that have worked in our history. Phrases that have slowed the thread or caused friction.

4. The Opening Line
One line. Fits their reading style based on the past messages.

5. My Trap
One paragraph. The mistake I am most likely to make with this specific client, named with a past message that shows the pattern.

Follow-up modes.
If I type "Higher stakes," rewrite the full brief treating the conversation as if the relationship ends if it goes wrong. Sharpen the concerns and the trap.
If I type "Lower stakes," rewrite the full brief treating the conversation as recoverable. Soften the opening line.
If I type "They have more power," rewrite the full brief assuming the client has the leverage.
If I type "I have more power," rewrite the full brief assuming I have the leverage and warn me against abusing it.

Rules.
Pull every pattern from the NotebookLM source. Never invent a quote from our messages. Never assume a behavior that does not appear in the history.
Never assume bad faith. Never give a script to read out loud. Give frames I can use in my own words.
If the conversation type has no precedent in our messages, say so in one line at the top of the brief and proceed with the closest adjacent pattern.
Never use the words "navigate," "essential," "crucial," "ensure," or "thoughtful."

Click “+” inside Knowledge. Pick the NotebookLM notebook you just created. Click “add”.

Fill the name of this gem, the description and so on.

The assistant is ready. Let’s test it.

The client did not answer my messages for 5 days. Answer without referencing any person or project name.

Here is the output.

If a client has gone quiet on you, this gives you the frame to break the silence using their own patterns, not generic communication advice.

Why not just ask ChatGPT how to have a hard conversation?

Try it. ChatGPT will give you a list of generic tips. Stay calm. Use “I” statements. Don’t be defensive. The advice will be true and useless because it does not know this client.

NotebookLM read 1.5 years of your actual messages with this person. It knows how they hesitate, how they push back, how they say yes. Gems keeps that knowledge fixed across every conversation. ChatGPT knows none of this.

ChatGPT gives you a tip sheet.

This stack gives you a strategy built from your real history with one human.

Subscribe now

Workflow 4: Personal Finance Decision Maker (Gemini)

My old car kept breaking down.

Oil leaks. Electrical faults. Something new every month.

My wife kept telling me to fix it myself. You are a mechanical engineer, she said.

Yes, I am. But my thesis was on gas turbines used in aircraft engines. I work on jet propulsion. What do I know about an Opel engine?

So we sat down one night and made the call.

We were not fixing this car anymore. We were buying an electric one. The kind I had run thermodynamic cycle calculations on in university but never owned.

That meant a loan. A real one. Monthly payments stacked on top of everything else we already pay.

I did the math before signing. Everything looked fine. A couple of months later I could not pay the full balance on my credit card. So I went back and looked at what was actually going on.

Layer 1

Go to NotebookLM, click “Try”, then “Create new notebook.”

Upload your expense data. The last 6 to 12 months works best.

Bank statement exports as PDF or CSV.

Credit card statements.

Any subscription tracker you keep.

Income exports from your main sources.

Click on customization, then custom. Paste this prompt.

You are my personal finance analyst.
You read only from the expense and income data I uploaded.
For every question I ask, return:
1. The top 5 expense categories from my data, with the monthly average and the trend over the period I uploaded.
2. The recurring fixed costs vs the variable costs, separated.
3. The biggest single expense category I am probably underestimating.
4. The monthly free cash I have left after fixed costs.
5. The one number I should know before I make any new financial commitment.
You never round to make a story. 
You show the math. 
You never recommend a purchase or a financial product.

Save it.

Let’s test it. I asked NotebookLM what my top expenses were.

Two of the five surprised me. I knew rent and groceries were big.

The third one I had been underestimating for months.

That is what I needed to see.

But I wanted to see it as a chart, not read it as a paragraph.

And I wanted to keep asking follow-up questions in the same place.

So I built the Gemini app.

Layer 2

Visit Gemini. Click “+”, then NotebookLM. Connect the notebook you just created.

Click Canvas in the tools menu next to “+”.

Paste this prompt.

You are building me a personal finance app inside Canvas. The app uses the NotebookLM source I connected. That source holds my expense and income data.

Build the layout with two tabs.

Tab 1. Dashboard.
A visual breakdown of my finances drawn from the NotebookLM source.

Top of the tab. A bar chart of my top 5 expense categories with the monthly average for each. Color the largest one in red.

Middle of the tab. A pie chart splitting fixed costs vs variable costs. Show the percentage and the actual amount.

Bottom of the tab. Three large numbers. Monthly income average. Monthly fixed costs total. Monthly free cash left.

No tables. No paragraphs. Just the charts and the three numbers.

Tab 2. Assistant.
A chat interface. I ask questions, the assistant answers using only the NotebookLM source.

When I ask a question about my finances, return the answer in this format.

The number first. One line.
The math behind it. Two sentences max.
What this means for a new financial commitment. One sentence.

Rules for both tabs.
Pull every number from the NotebookLM source. Never invent a value. Never guess at a category I did not upload.
Never recommend a purchase, a loan, an investment, or a financial product. If I ask for one, return the math that the decision depends on instead.
Show the math when I ask for it. Never round to a story.
If a category in my data is unclear, ask me which bucket to map it to before answering.
Never use the words "essential," "crucial," "ensure," "navigate," or "smart" when talking about money.

This is how my screen looks like.

Hit enter. Two minutes later, the app is ready.

I clicked share, copied the link, and opened it in a new browser.

The dashboard showed the bar chart, the fixed-versus-variable split, and the three big numbers. The assistant tab let me ask a question and get back the number, the math, and what it means for a new commitment.

I asked it the question I had been avoiding for two weeks. Can I afford the car payment without changing my lifestyle? It returned the number. Then the line item I would have to cut to keep my current savings rate.

The math was tight. The car payment stayed.

Subscribe now

Why not just upload your bank statements to ChatGPT?

Try it. You’ll get a paragraph telling you to budget better.

And what if it hallucinates? You can tell your bank this:

“Sorry, sir, the AI hallucinated, so can I skip my payment this month?”

They’ll laugh a lot, but very briefly.

What’s Next

Four workflows. One stack. Real builds.

The setup prompts above are the whole scaffolding.

Copy them. Replace my sources with yours.

The point is not that NotebookLM is magic.

The point is that NotebookLM + Gems + Gemini Canvas + Claude is the cleanest stack on the internet for turning your real data into something that thinks for you on demand.

If you build one, reply to my welcome email. I read everyone.

If you want to keep going, here is where I would go next.

• If you want to see how deep one workflow can go. I built a self-updating trading bot that runs on a loop. NotebookLM analyzes the trade journal every Sunday, Claude rewrites the strategy, and OpenClaw deploys it. Hit 77% win rate. Read I Built a Self-Updating Trading Bot (Claude & NotebookLM & Karpathy’s Method).

• If you want NotebookLM at the master level. Same stack, deeper prompts, the full feature map. Most people use 10% of what NotebookLM can do. This post covers the other 90%. Read How to use NotebookLM better than 99% of People.

• If you want to build apps, not just assistants. The pipeline I used to build a $10M CRM dashboard in 8 minutes. NotebookLM distills, Gemini prototypes, Claude Code builds. Read How to Build Apps with NotebookLM Better Than 99% of People.

• If you want my 10 best NotebookLM prompts for studying. The post that drives most of my SEO traffic. Mind maps, flashcards, practice exams, dependency-ordered learning paths. Read 10 NotebookLM Prompts For Studying (Beat 99% of Students).

• If you want 10 more NotebookLM prompts, but for everything else. The next 10 prompts in the series. Build apps, generate reports, automate research, run agent orchestration. Read 10 NotebookLM Prompts That Put You Ahead of 99% of People.

• If you want the full agent stack with Claude Code. I built a system where Agent 1 reads meeting notes, Agent 2 researches the topics, Agent 3 writes the report. All three use NotebookLM. Read How I Built a Multi-Agent Voting System Using NotebookLM and Claude Code.

• If you want Claude to remember your work across sessions. My Obsidian + Claude second brain reads my vault and answers client messages when my brain stops at 2 PM. Read My Brain Dies at 2PM. I Built an Obsidian + Claude Second Brain.

• If you want to ship apps with Claude Code. The Build. Ship. Repeat. series. Seven months. Seven apps. One in public. Read Claude Code: Build. Ship. Repeat..

The full application prompts for every workflow above, plus 245 more NotebookLM prompts I have tested across health, career, finance, law, and relationships, are in the Vault.

What is the vault?

If you go paid(annual or inner circle), you get all of them, plus everything I ship next.

See you there.

Calendar, Gmail, Slack, Notion, and my Claude Routine doc, in that order, every day.

The tabs weren’t the problem.

Nothing talked to anything else.

Calendar said one thing, Gmail another, Slack a third.

Ten minutes had gone by before I picked what to work on.

And I was okay with it, until Claude made our lives easier with this update with Claude's cowork live artifacts.

Now my morning routine looks like this.

Two things to add.

You can add your Claude routines here.

Also, when I feel lazy, I just open the assistant and let it explain everything by voice.

How Claude Live Artifacts Works

An artifact in Claude lets you build AI apps.

In my post analyzing 11 Substack creators' Notes, I built one artifact per creator.

Here’s mine.

A Live Artifact is different. Every time it loads, it can reach out to your connected applications and pull fresh data.

Gmail, Calendar, Slack, or whatever else you've connected.

It can also call Claude directly from inside its own code. That’s how the voice assistant works in the corner.

One more thing I discovered. It can connect to your scheduled routines and surface their output. That's how I tracked Claude's news.

The architecture bends to your workflow. We'll build one now.

A Note Before We Build

This is the fourth post of Claude. Build It.

Each post will feature how you are running the same use case in your own environment. Different decisions, different stacks, different connectors.

DM me how you built yours. I’ll feature your version in the next post, like Dan’s.

Let’s build it together.

Community Spotlight: Dan’s PR Reviewer

Where LearnAIWithMe community members show how they built it.

Welcome to Build it #3.

Each case stands alone. Different problem, same shape of solution.

In Build it #2, I built a health Advisor for my pregnant wife, which queries PubMed.

Dan saw this and rebuilt it for code reviews.

He uses the same skeleton, but different organs. See his PR Reviewer structure.

Dan publishes open-source tools for AI agents.

Most of his code is written by Claude or Codex.

AI writes the PR. AI reviews the PR. AI says everything is fine.

That is the same sycophancy problem from Build It #2.

So Dan replaced the reviewer:

• Gate 1. Tests run on every PR. Pass means move on. Fail means stop.

• Gate 2. A fix agent (Codex with GPT 5.5) looks only at what failed and what changed. Small patches, no rewrites.

• Gate 3. If the fix isn’t confident, or the PR touches sensitive code, Dan gets a Telegram ping. No merge until he checks.

Dan’s note:

What I borrowed from Gencay’s Build It #2 was the shape of the process, because he didn’t build a Claude agent that tries to sound certain for his pregnant wife, he built one that checks PubMed before answering, cites the research when it finds a clear answer, and admits when the literature is unclear.
I adapted that same idea for PRs, so tests run first, Codex only gets scoped to the failing area, and when it can’t prove the fix, it sends me a Telegram note with what it tried and where it got stuck.

Dan also has a Substack called Vibe Stack Lab.

You can follow him here;

Vibe Stack Lab

Practical guides to AI agents, useful models, and real-world workflows that help you get better results from AI.

By Dan Cucolea

Morning Briefer With Claude Cowork Live Artifact

Four steps, and it takes about fifteen minutes if your connectors are already authenticated. You'll need Cowork access.

Read more

Four weeks to go, and the questions are never-ending.

Is this normal? Should I be worried? Why does this hurt now and not yesterday?

Most of these questions do not need a doctor.

They need a fast, honest answer at 2 AM. The doctor’s office is closed.

The pregnancy book on the shelf was written in 2019.

Google sends you to a forum where someone says they had the same symptom and it turned out to be nothing, and below it someone else says it turned out to be everything.

So you open Claude. Or ChatGPT. You type the symptom. You get an answer.

That is the moment I stopped.

Community Spotlight: David's Marketing Council

Where LearnAIWithMe community members show how they built it.

On Monday, I asked you to build your own Council.

David Forer did.

He picked Hormozi, Gary V, and Seth Godin.

A modern marketing trio. Then he asked them:

"Should I consult or productize my AI operations?"

Three rounds. Nine positions. One slide deck.

David's note:

It surprised me just how accurate the results are. I would watch it building (click the down arrow in Claude) and see each response, and I was like “Yep, I could hear Gary saying that”.

If you read the whole PDF, you could tell who was behind each slide, which shows the effectiveness of this council.

There are really so many applications of this, but with the right question, you could do a video, a Twitter post, or a LinkedIn carousel, among other things.

My next step is expanding the books used by each marketer.

Hormozi has three big 100M books, Gary Vee has 2 more that will help, and Seth has many books. The goal would be to get even deeper answers from each author and create even more tension.

And this is the kind of build that makes me regret not starting this series sooner.

Look closer at what David said, once again.

My next step is expanding the books used by each marketer.

I love it, so smart and effective.

Here is David’s website, if you want to keep in touch with him.

Join the “Build it” like David

David already built his.

So you can.

I said this in Build It #1: each post features how you’re running the same use case in your own environment.

In case you did not know, we started a new series “Build It”, where we’ll build 30 builds in 10 weeks.

If you run yours this week, you're in the next one.

DM me how yours turned out.

If you get errors, DM me so we’ll solve together.

Let me include you in the next post!

Let's Build it together.

The data that scared me

Anthropic published a study on April 30, 2026.

They sampled 1 million claude.ai conversations and looked at how often people ask the model what to do with their lives.

Health and wellness was the single biggest category at 27%.

More than career. More than relationships. More than money.

People are asking AI about their bodies more than anything else.

Across all guidance conversations, Claude was sycophantic 9% of the time.

Agreed too easily. Reassured without challenging.

Told the user what the user wanted to hear.

Now do the math.

If 27% of guidance is health, and 9% of all guidance is sycophantic, the absolute number of sycophantic health conversations is large.

Anthropic itself flagged the pattern.

The model is more likely to agree under pressure.

People asking about their bodies push back when they do not get the answer they want.

That is the failure mode I cannot afford.

So I built a different kind of agent.

What we'll build: a Claude Code agent

It is not a doctor. It is not a diagnostic tool. It is a literature filter.

My wife asks a question on her phone.

The agent does not answer from the training data. It calls PubMed first.

It reads recent peer-reviewed papers. It returns what the literature says, with citations.

If the literature is unclear, the agent says the literature is unclear.

If the symptom is in the “call your doctor now” category, the agent says that and stops.

The agent never replaces her OB.

The agent is what she reads at 2 AM when the OB is asleep, and the question is, “is this normal or am I overreacting?”

It is what most people would search Google for. It just searches PubMed instead, and reads the papers, and tells her what the actual evidence says.

The architecture: 3 checks, 1 Claude Code agent

Three checks. One agent. One always-on machine.

Urgency check runs in the Claude Code agent. The system prompt holds the red-flag list. Heavy bleeding, severe pain, signs of preeclampsia, decreased fetal movement after 28 weeks, the agent stops, says “Call your OB now,” and ends the session.

PubMed query goes straight to PubMed. Last 5 years, human studies, English. The agent reads the top papers itself. No third-party scraper.

Cite or stop is enforced by the system prompt. Every claim ships with author, year, and journal. If PubMed has nothing, the agent says so. No fallback to training data.

All of it runs on a machine that stays on 24/7.

I use a Mac mini. If you don’t have one, leave your PC or laptop on with Claude Desktop open, same effect.

Dispatch routes her phone questions in. The machine waits. The agent answers in under a minute.

Read more

TL;DR: In Part 1 you connected Claude to your real data and watched it build working tools inside the chat. Part 2 is where it gets personal. Skills teach Claude how you work, so every new chat starts knowing your company, your voice, and your standards.

I built a Skill last month for client emails. Three writing samples, five banned phrases, and one paragraph on tone.

I have not written a client email from scratch since.

That is a Skill doing its job.

Part 1 of this series ran on Leadership in Change. I covered Connectors and Artifacts, the two features that change what Claude can see and what Claude can build.

Part 2 is Skills. And I handed it to Joel Salinas, because he uses them really sharply. Joel writes Leadership in Change, the newsletter for leaders who want to use AI without the hype.

Leadership in Change

Don’t just read about AI, lead it. Get the weekly playbook for leaders who want to supercharge their expertise and multiply their impact.

By Joel Salinas

Over to Joel.

— Gencay

One sentence into a fresh chat, Claude already knows my company, my audience, my brand voice, my standard offers, and the three things I never want it to write. I haven’t typed a single line of context. That’s a Skill doing its job.

In Part 1, Gencay walked you through Connectors and Artifacts. If you sat in Claude for twenty minutes and actually built that prospecting workflow he showed you, you already feel the shift. Claude can read your Drive. Claude can build a working tool inside the chat. That alone changes the shape of your week.

If you missed Part 1, go read it first. It’s the foundation: Learn 80% of Claude in Under 20 Minutes — Part 1.

Now I want to take you one layer deeper, because there’s a feature inside Claude that quietly does more for me than Connectors and Artifacts combined.

It’s called Skills. And once you set one up, you stop opening Claude like a chatbot and start opening it like a colleague who already knows your company.

What is a Claude Skill?

In one sentence: a Skill is a small set of instructions, examples, and reference files that Claude loads automatically when it’s relevant, so every new chat starts pre-briefed instead of blank.

Think of it this way. Connectors give Claude ingredients (your emails, your docs, your data). Artifacts give Claude a kitchen (a working interactive tool). Skills give Claude the recipe (the way you cook).

Without a recipe, you have ingredients and a kitchen and a generic meal. With a recipe, the same ingredients and the same kitchen turn into your exact dish, the way you make it, every time.

The leaders I coach who use Claude for thirty minutes a day and the ones who use it for thirty hours a day are using the same software. The difference is almost always Skills.

How to set it up

You build Skills inside Claude Cowork, the workspace inside Claude where you can browse and create your Skills directly. The whole flow:

1. Open Claude Cowork.

2. Tell Claude you want to create a new Skill, and describe what you want it to do, in plain English.

3. Claude drafts the Skill in front of you.

4. Click Save Skill.

That’s the entire lift. There is no terminal. There is no API key. There is no markdown file you have to write or move into a folder. You describe the Skill the way you’d brief a new team member, and Claude turns that brief into a saved Skill that loads automatically the next time it’s relevant.

Use Case: Beginner — Build Your First Skill in 10 Minutes

Here’s the flow I’d walk you through if you opened Claude right now and told me you’d never built a Skill before.

Pick the writing task you do most often. Could be client emails. Could be project status updates. Could be the weekly note you send your team. Whatever it is, you have an opinion about how it should sound, and you’ve corrected Claude on it more than once.

Open Claude Cowork, and paste something like this:

Prompt: Create a new Skill called “[your-skill-name].” 
It should help me draft [the task you do most]. 

Voice rules: [3 to 5 things you always want]. 

Things to avoid: [3 things you never want]. 

Audience: [who reads it]. 

Standard format: [length, structure, sign-off].

Claude will ask you a few clarifying questions. Answer them like you’re onboarding a new hire. Then Claude drafts the full Skill in front of you. When it looks right, click Save Skill.

Open a new chat. Ask for the thing. Watch Claude do it correctly the first time, in your voice, in your format, without a single line of context from you.

That’s the moment leaders stop being skeptical about AI and start asking what else they can hand off.

Use Case: Advanced — Skills + Connectors + Artifacts, Stacked

Now we put all three pieces together. This is the workflow that pays for itself in a week.

Remember Gencay’s Part 1 demo, where Claude opened a Drive doc, ran each company through Vibe Prospecting, and handed back a ranked outreach list? Powerful, and the output was clean. But the next step was still on you. You had to write the actual outreach.

Watch what happens when a Skill is loaded on top.

Same Drive doc. Same Connectors. But this time, before you ask for the list, you have a Skill called outreach-voice already loaded with your tone, your standard offer, your three opening lines, and the three openers you never want to use again.

Prompt: Open my Drive doc “Q2 Target Companies.” 
For each company, use Vibe Prospecting to pull funding, 
tech stack, and hiring trends. 

Flag the ones that raised in the last 6 months or are 
hiring for roles related to our service. 

For the top 5, draft a personalized outreach email in my voice 
using the outreach-voice Skill. Show the drafts inside an 
Artifact I can edit and copy from.

One prompt. Three layers running together.

• Connectors pull the company list and enrich each one with live data.

• The Skill writes each email in your tone, with your offer, in your format.

• Artifacts present the five drafts in an editable card you click through and copy.

You went from “build me a list” to “build me a list, research them, write the outreach in my voice, and stage it for me to send.” End-to-end. One chat. Twenty minutes of leader-time replaced with twenty seconds of leader-attention.

This is when the question stops being “should we be using AI?” and starts being “why is half my team still doing this manually?”

Top 5 Skills Every Leader Should Build First

If you’re not sure where to start, build these five. In this order. They cover ~80% of where Skills create real time savings.

• Brand Voice Skill — Drops your tone, audience, banned words, and three writing samples in. Every draft sounds like you.

• Meeting Prep Skill — Tells Claude how you walk into meetings. The questions you always ask. The risks you always check.

• Proposal Reviewer Skill — Scores any proposal sent your way on clarity, feasibility, and ROI. Flags vague language. Suggests three questions before you sign.

• Decision Framework Skill — Loads the mental models your team actually uses (Eisenhower, RICE, whatever you run). Now Claude scores trade-offs the way you score them.

• Client Profile Skill — One Skill per top client, loaded with their company, their goals, their constraints, and the project history. Every email to that client starts pre-briefed.

These five together will save you more time in the first month than every productivity app you’ve installed this year.

A practical note on quality

The Skill is only as sharp as the context you put in it. A vague Skill (”write in a friendly, professional tone”) gets you a vague output. A Skill with three real writing samples, five banned phrases, and a named audience gets you the draft you actually wanted on the first try.

Your first version of any Skill will be okay. You’ll use it for a week, notice three things it gets wrong, and update it. By version three, it reads like Claude is finishing your sentences. Plan for the iteration. It’s the work that makes the rest of the work disappear.

What’s next?

That’s Part 2. You now have Connectors, Artifacts, and Skills. Most leaders I coach who get all three pieces stood up tell me the same thing about a month in: they cannot remember how they used to work without it.

If you want to go deeper on the leadership side of this, on how to actually run a team that uses AI well without losing the human part, that’s what I write about every week at Leadership in Change. We cover the strategy, the org design, the trust questions, and the practical setups like the one in this post. Subscribe at leadershipinchange.com.

Leadership in Change

Don’t just read about AI, lead it. Get the weekly playbook for leaders who want to supercharge their expertise and multiply their impact.

By Joel Salinas

And subscribe to Gencay here at LearnAIWithMe for the next round of practical Claude builds. He’s the no-code teacher I wish I’d had when I started.

A quick personal note on why this matters. I have a two-year-old and a seven-year-old at home. Every hour I get back from busywork is an hour with them, or an hour on the deeper work that actually moves my businesses forward. Claude Skills are quietly the biggest reason I have those hours. Build one today. Pick the task you do most. Write down how you do it. Save it. Tomorrow Claude already knows.

That is the moment Claude stops being software you use and starts being part of how you work.

Written by a human, for humans.

Joel

Most of the time, I find myself thinking about the geniuses I look up to. What would they do?

• Nikola Tesla was one of the greatest inventors who ever lived. A mind ahead of his century, no doubt.

• Marcus Aurelius, who could decide calmly even in storms. A Stoic. The man who put his own actions at the center of everything.

• Sun Tzu, who knew how to win without fighting.

But which one? Each shows up for me at different moments.

• Tesla told me to go deeper when I wanted to reach more people on Substack,

  • He would have written a series, AI said once. So I started Build It and Build. Ship. Repeat.

• After a great client meeting that went into radio silence, and the anxiety that followed, Marcus Aurelius told me the only thing in this life I can change is my own actions. Only that.

• When I wanted to argue with the problematic neighbor whose loud grill smoke was filling our flat, Sun Tzu’s line came to mind: “The supreme art of war is to subdue the enemy without fighting.”

The pattern was clear. I was already running a one-person council. Just badly.

I was rereading the same passages.

Quoting from memory.

Half-remembering chapters.

So I built the Council, using Claude Code Agents. (With one prompt)

Three voices, three books, three rounds of argument, one report at the end.

NotebookLM read the books. It runs the rounds. It builds the Slides, consisting of each voice.

I just read the friction.

The day they argue, I decide.

Not before.

And we’ll do it in 3 steps, together.

Let’s start.

A note before we build

This is the first post of Claude: Build It.

Each post, I hope, will feature how you are running the same use case in your own environment. Different decisions, different books, different stacks.

DM me how you built yours. I’ll feature your version in the next post.

Let’s build it together.

How does the system work?

First, Claude Code spins up four agents: three voices and one Reporter.

Each voice (Sun Tzu, Marcus Aurelius, Tesla) queries NotebookLM through the notebooklm-py CLI. Every query hits only its own book. No drift. No blending.

They argue in three rounds. Round 1: initial positions. Round 2: cross-examination. Round 3: final stance.

The Reporter listens to all three. Compiles the transcript. Name the friction.

Then the Reporter sends the deliberation back to NotebookLM. NotebookLM builds a 12-slide deck, one slide per voice per round, plus the friction and the final question.

You read the friction. You decide.

What will the end report look like?

The prompt is yours to adjust. You change the decision. You change the books.

You paste your decision. The system spins up three agents, each one sourced from a single book, each one queried only against its own author.

They argue in three rounds. The Reporter listens, compiles, and names the friction. Then NotebookLM turns the whole deliberation into a 12-slide deck with sourced quotes on every slide.

The final report looks like the slide below.

Each voice gets its own slides. The General lays out his terrain. The Stoic names what is in your control. The Inventor draws the future. Each slide cites the chapter the quote came from.

Three voices. One report. A decision that has been argued before it reaches you.

The Council waits. Build it.

Read more

Six months later they are still on YouTube, scrolling tutorials.

They know what RAG stands for but they cannot build anything.

The problem is not effort. They do not know how to learn AI from zero.

AI is not a topic. It is a stack. Concepts on top of concepts.

This is where most beginners quit.

NotebookLM fixes this.

You upload the sources. You prompt it the right way. It becomes a tutor that meets you at zero.

So let’s discover the ten best notebookLM prompts for learning ai.

All tested on Anthropic’s “AI Prompt Engineering: A Deep Dive” youtube video.

Copy-paste ready.

If you are starting from zero, these will replace half the apps in your learning stack.

10 NotebookLM Prompts for Learning AI

NotebookLM has 8 features.

Most beginners use one. Audio Overview. They click generate. They get a generic podcast.

They close the tab.

That is not learning. That is downloading.

Real learning happens when you customize.

The prompt box on every feature is where the work happens.

Type the right thing there, NotebookLM stops being a search engine and starts being a tutor.

Here is what we will cover.

Each prompt is paired with the NotebookLM feature that complements it.

Each one has a fill-in template.

Each one is tested on Anthropic’s “AI Prompt Engineering: A Deep Dive” youtube video.

Zero coding. Nothing to install. Everything happens inside NotebookLM.

Let’s go.

How to train NotebookLM?

Before starting, let me show you how I trained the notebookLM, using Anthropic’s “AI Prompt Engineering: A Deep Dive” YouTube video.

Steps: Copy the YT link - Visit NotebookLM - Click on Try - Click on Create New - Select websites- Paste the link

1. Onboarding Prompt That Explains AI Like You Are 10 Years Old

NotebookLM Feature: Report (Create your own)

UI selections: Format → Create your own. Length → Default.

The first wall when you learn AI from zero is jargon.

RAG. Tokens. Embeddings. Fine-tuning. Chain of thought.

Most people ask ChatGPT “what is RAG”. They get a textbook definition.

Two hours later they still cannot explain it to a friend.

The problem is not the definition. The problem is the layer.

A 10-year-old explanation strips jargon. An adult explanation gives the working knowledge. A practitioner layer shows the trade-off. Three layers. One concept. Each earns its place.

This is one of the best NotebookLM prompts for learning AI because it forces the model to do the work most beginners skip.

Steps: Open your notebook → In Studio, click Reports → Choose “Create your own”.

For every key concept in the sources, write a three-layer explanation.

Layer 1. Explain it to a 10-year-old. One sentence. Use a real-world 
analogy. Zero jargon. If a term needs another term to be understood, 
break it down further.

Layer 2. Explain it to a curious adult learning AI from zero. Three to 
four sentences. Define one technical term per concept, but only after the 
analogy makes the idea click first.

Layer 3. Explain it to a junior engineer who needs to use it tomorrow. 
What is the trade-off. When does it break. What does the documentation 
forget to mention.

Order concepts by dependency, not by source order. Concepts that depend 
on other concepts come after their dependencies. Mark each concept with 
the prerequisites needed to understand it.

If a source contradicts another source on a concept, flag it in Layer 3 
and explain which definition is more current.

I tested this on the Anthropic notebook. Here is one concept it produced.

Three layers. One concept. One pass.

If the Layer 1 analogy does not click, the model did not understand the concept either. Re-prompt with a source quote.

Run this once on a new notebook, and you get a complete glossary, ordered by dependency, ready to study.

Subscribe now

2. Audio Overview That Debates Two AI Experts on the Same Topic

NotebookLM Feature: Audio Overview

UI selections: Format → Debate. Length → Default. Language → English.

The default Audio Overview gives you two hosts agreeing with each other for 15 minutes.

Calm voices. Clear summary. Zero friction.

That is the problem.

When you learn AI from zero, you need to hear where experts disagree.

RAG vs fine-tuning. Pretrained vs RLHF.

Prompt engineering vs prompt programming.

The disagreement is where the real understanding lives.

This prompt turns the Audio Overview into a structured debate.

One host defends the textbook view. The other host challenges it with the practitioner view.

You learn both sides in one listen.

Steps: Open your notebook → In Studio, click Audio Overview → Choose Debate format.

Structure the conversation as a debate between two AI practitioners.

Host A is a researcher who learned AI from textbooks and papers. Defends 
the standard definitions. Cites sources directly.

Host B is a builder who learned AI by shipping products. Challenges 
textbook claims with real-world trade-offs. Cites edge cases where the 
standard definition breaks.

For every key concept in the sources, run this pattern. Host A defines 
the concept in plain language. Host B agrees on the definition but 
challenges one assumption. Host A defends with a citation. Host B counters 
with a scenario where the definition fails.

End each concept with a one-sentence synthesis both hosts agree on.

Avoid generic agreement. If Host B has nothing to challenge, skip the 
concept. Force tension on every topic that gets airtime.

I tested this on the Anthropic notebook. Here is what came out.

Two angles. One concept. Real tension.

The trick is the synthesis line. It forces the model to commit to a takeaway instead of leaving you with two opinions and no decision.

Listen to this on a walk.

By the end you understand the concept better than after reading three articles.

Subscribe now

3. Mind Map Prompt That Builds the AI Stack From Zero to Advanced

NotebookLM Feature: Report (Create your own)

UI selections: Format → Create your own. Length → Default.

When you learn AI from zero, the biggest trap is order.

You watch a tutorial on RAG before you understand what an embedding is.

You read about fine-tuning before you know what pretraining did.

Each video assumes you already learned the prerequisite.

NotebookLM has a built-in Mind Map.

It shows hierarchy. It does not show dependency.

Hierarchy answers “what are the categories”.

Dependency answers “what do I learn first”.

For a beginner, the second question is the only one that matters.

This prompt builds a dependency-ordered learning path from your sources.

Steps: Open your notebook → In Studio, click Reports →Choose “Create your own”.

Build an AI learning stack from the sources. Order it by dependency, not 
by topic.

Output structure.

Level 0 - Foundations. List concepts that depend on nothing else in the 
sources. These are the entry points. One sentence per concept.

Level 1 - Built on foundations. List concepts that need exactly one 
Level 0 concept to make sense. For each, name the prerequisite.

Level 2 - Composite. List concepts that combine two or more Level 1 
concepts. For each, name both prerequisites and explain the combination 
in one line.

Level 3 - Frontier. List concepts the sources mention but do not fully 
explain. These are what to learn next, outside the current notebook.

For every concept, mark whether the sources teach it directly, mention 
it in passing, or assume the reader already knows it. Flag the assumed 
ones. Those are the gaps.

End with a recommended study order. Number each concept 1 through N.

I tested this on the Anthropic notebook. Here is the output.

Now you have a path. Not a list of topics. A path.

The “gaps” section is the most useful part. Those are the prerequisite concepts the sources skipped, which is why beginners feel lost halfway through any AI tutorial.

Run this once, and you stop watching tutorials in the wrong order.

Subscribe now

4. Flash Cards That Test Scenarios, Not Definitions

NotebookLM Feature: Flash Card

UI selections: Number → More. Difficulty → Hard.

NotebookLM’s default flash cards work like a textbook index.

Q: “What is RAG?”

A: “Retrieval-augmented generation.”

You memorize the answer. You still cannot use the concept.

This is the same trap Amanda Askell describes in the Anthropic prompt engineering roundtable. Real testing is not “can you recall the definition”. Real testing is “can you pick the right tool when the situation is messy”.

When you learn AI from zero, the gap between knowing a term and using it is where most beginners get stuck. Scenario cards close that gap.

Steps: Open your notebook → In Studio, click Flashcards.

Generate flash cards that test decision-making, not recall.

Each card has two parts.

Front. A specific scenario where the user must choose between two or 
more concepts from the sources. Include realistic constraints. Token 
limits, latency, dataset size, cost, accuracy requirements. Make the 
choice non-obvious.

Back. The correct choice. A one-sentence reason grounded in the sources. 
A note on what the wrong choice would cost in this scenario.

Cover these scenario types.

1. Tool selection. User wants outcome X. Pick between technique A and B.
2. Failure diagnosis. Output looks wrong. Pick the most likely cause 
from two or three candidates.
3. Trade-off. Pick the option that fits the constraint, not the option 
that sounds best in isolation.

Skip cards where the answer is in the question. Skip cards that test 
vocabulary. If a concept only has a definition in the sources and no 
applied example, do not card it.

I tested this on the Anthropic notebook. Here are the flashcards.

This card forces you to balance four constraints.

You cannot answer it by recalling a definition.

You answer it by reasoning about the scenario.

If you can answer 20 cards like this, you understand RAG.

If you can only answer “what does RAG stand for”, you do not.

Run this on every new notebook. Skip the default cards.

Subscribe now

5. Quiz Prompt That Catches What You Almost Understood

NotebookLM Feature: Quiz

UI selections: Number → More. Difficulty → Hard.

When you learn AI from zero, the most dangerous state is “I think I get it”. You read the chapter. You nodded along. You can summarize it back. But you cannot solve a problem with it.

Amanda Askell calls this the “almost understood” zone. The model passes a generic quiz but breaks on edge cases. The same is true for you.

Default quizzes test the happy path. “What does RAG stand for?” If you can answer that, the quiz says you understood. You did not. You just memorized.

This prompt builds quiz questions that target the exact spots where almost-understanding hides.

Steps: Open your notebook → In Studio, click Quiz.

Generate questions that target partial understanding.

For each concept in the sources, build one question that follows this 
pattern.

The question describes a situation where the concept applies. The 
correct answer identifies the specific mechanism the sources describe. 
Wrong or tempting answers must come from these confusion sources.

Confusion 1. A neighboring concept that does something similar but works 
differently.
Confusion 2. A common misconception the sources explicitly correct.
Confusion 3. An oversimplification that drops the key detail.

After each question, include a one-sentence explanation for why the 
common confusion is tempting and what specific detail rules it out.

Skip questions that can be answered by recognizing keywords. The 
question must require the user to think about which concept fits the 
situation, not which word matches the definition.

I tested this on the Anthropic notebook. Here is the quiz it produced.

This question is hard not because it is obscure. It is hard because every wrong answer sounds reasonable to someone who almost understood chain of thought.

If you can answer all questions like this without checking notes, you actually understand the concept. If you skip the explanations, you do not.

Run this after every study session. The wrong answers tell you what you almost knew.

Subscribe now

6. Slide Deck for Teaching Back What You Learned

NotebookLM Feature: Slide Deck

UI selections: Length → Default.

When you learn AI from zero, the fastest test of real understanding is teaching it back. Richard Feynman’s rule.

If you cannot explain a concept on a slide, you do not understand it yet.

Most beginners skip this step.

They read, watch, take notes, then move to the next topic.

The concept fades in two days.

A teach-back deck forces structure. One concept per slide.

The slide must be self-contained, someone who never read the source must understand it from the slide alone.

If you cannot fit the concept on one slide without losing the meaning, you have not compressed it correctly. That gap shows you what to re-study.

Steps: Open your notebook → In Studio, click Slide Deck.

Generate a teach-back deck where each slide explains one concept to a 
peer who has not read the sources.

For every slide, follow this exact structure.

Slide title. The concept name. No subtitle. No clever phrasing.

Three bullets. Each bullet is one short sentence. Maximum 10 words. 
Bullet 1 is the definition in plain language. Bullet 2 is the mechanism 
or reason it works. Bullet 3 is the most common misuse or limitation.

Speaker notes. Three sentences maximum. Sentence 1 is the analogy that 
makes the concept click. Sentence 2 is the example from the sources. 
Sentence 3 is the question a confused listener would ask, and the 
one-line answer to it.

Skip concepts that cannot be compressed to this structure. Mark them as 
"needs more study" at the end of the deck instead of forcing them onto 
a slide.

Order the slides by dependency. Foundations first. Composite concepts 
last.

I tested this on the Anthropic notebook. Here is one slide it produced.

This slide passes the teach-back test.

A peer who never read the sources can understand chain of thought from this slide alone.

If a slide does not pass that test, the concept is not yet yours.

Run this after each study session. Present three slides to yourself out loud. The ones you stumble on are the ones to re-study.

Subscribe now

7. Report Prompt That Finds the Gaps in Your Understanding

NotebookLM Feature: Report (Create your own)

UI selections: Format → Create your own. Length → Default.

When you learn AI from zero, the hardest part is not learning. It is knowing what you do not know. You read a chapter, feel confident, then hit a question that exposes a gap you did not see coming.

Amanda Askell described the same problem in the Anthropic roundtable. Before she trusts a prompt, she asks the model to flag where the instructions are unclear. The same trick works for your own learning.

This prompt turns NotebookLM into a gap detector. It reads the sources and tells you what the sources assume you already know, where they are vague, and where they contradict each other. Those three categories are the exact spots where beginners get stuck without realizing it.

Steps: Open your notebook → In Studio, click Reports → Choose “Create your own”.

Audit the sources for gaps in coverage. Output three sections.

Section 1. Assumed knowledge.

List every concept the sources reference but do not explain. For each, 
quote the sentence where the concept appears, and write one line 
explaining what a beginner would need to know to follow that sentence.

Section 2. Vague explanations.

List concepts the sources define but leave incomplete. For each, name 
the concept, quote the partial definition, and identify the specific 
question the definition fails to answer.

Section 3. Source contradictions.

List concepts where two or more sources disagree. For each, quote both 
sides, name the disagreement in one sentence, and flag which version 
appears more current based on the publication context.

Do not invent gaps. Only flag a gap if you can quote the source text 
that creates it.

End with a study recommendation. Rank the gaps by which one would block 
the most downstream concepts if left unfilled.

I tested this on the Anthropic notebook. Here is the trimmed output.

Now you have a list of what to study next. Not from a generic curriculum. From the gaps your specific sources left open.

This prompt prevents the most common beginner failure: studying in a loop without realizing the foundation is missing.

Run this on every new notebook before you start studying. The gaps come first.

Subscribe now

8. Video Overview for the Topic That Lost You in Class

NotebookLM Feature: Video Overview

UI selections: Format → Explainer → “Whiteboard” as a Visual Style

When you learn AI from zero, every topic has one moment where you lose the thread. The lecturer moves on. The textbook moves on. You are still stuck on slide 12 trying to figure out what an embedding actually is.

The default Video Overview gives you a generic walk-through of all the sources. Useful for an overview. Useless for the one concept that lost you.

This prompt turns the Video Overview into a targeted rescue mission. You name the exact concept that broke your understanding. The video focuses everything around that single concept — the prerequisites you need, the analogy that makes it click, the example that proves it, the misconception that probably tripped you.

One concept. One video. One unstuck moment.

Steps: Open your notebook → In Studio, click Video Overview → Choose Explainer format → Select “Whiteboard” as a Visual Style

Build the entire video around one concept: {concept_name}

Structure the narrative in five parts.

Part 1. The prerequisite check. Name the two or three concepts the 
viewer must already understand to follow this one. Define each in one 
sentence. If a prerequisite is missing from the sources, flag it 
explicitly so the viewer knows to study it first.

Part 2. The misconception. State the most common wrong way beginners 
understand this concept. Show why it is wrong using a quote from the 
sources.

Part 3. The mechanism. Explain how the concept actually works. Use a 
concrete analogy from everyday life before introducing any technical 
term. Visual theme should reinforce the analogy.

Part 4. The example. Walk through one specific example from the 
sources where this concept is applied. Show the input, the process, 
and the output.

Part 5. The trade-off. Explain when this concept is the right tool and 
when it is not. Name the alternative the sources mention and the 
condition where the alternative wins.

Do not summarize the other concepts in the sources. Stay locked on 
{concept_name}. If the sources do not cover one of the five parts, say 
so directly instead of filling space.

I tested this on the Anthropic notebook with {concept_name} set to Chain of Thought.

Here is the structure it produced.

The video stays on Chain of Thought for every minute of its runtime.

This is the difference between a video that summarizes the chapter and a video that unsticks the one part you missed.

Run this every time a concept loses you in class.

Name the concept in {concept_name} and let the video do the rest.

Subscribe now

9. Infographic That Turns Dense Theory Into One Screen

NotebookLM Feature: Infographics

UI selections: Visual Style → Bento Grid. Orientation → Portrait. Detail Level → Concise.

When you learn AI from zero, the worst content format is a wall of text explaining a process. You read paragraph 1, by paragraph 4 you forgot what paragraph 1 said. The information was there. The structure was missing.

An infographic forces structure. One concept becomes one screen with a visual hierarchy you can take in at a glance. Bento Grid layout breaks the screen into clean blocks, each block carries one piece of the concept, no block carries two.

This is not a poster. This is a study aid you actually use.

Steps: Open your notebook → In Studio, click Infographics → Choose Bento Grid.

Build a Bento Grid infographic for one concept from the sources: 
{concept_name}.

Output structure.

Headline block. The concept name and a one-line definition. Maximum 12 
words combined.

Block 1. The core mechanism. One sentence. One visual metaphor reference 
the designer can illustrate.

Block 2. One hard statistic or measurable fact from the sources. Quote 
the source phrasing. Maximum 8 words of context.

Block 3. The most common misuse. One sentence describing what beginners 
get wrong about this concept.

Block 4. The trade-off. Two short labels separated by "vs". Each label 
is the condition where one approach beats the other.

Block 5. The next concept. Name the concept that builds on this one and 
should be studied next.

Block 6. The prerequisite. Name the concept the viewer must already 
understand before this one makes sense.

Constraints. No block exceeds 8 words of body text. If the sources do 
not contain a hard statistic for Block 2, replace it with a direct 
quote from the sources that captures the concept in 8 words or less.

I tested this on the Anthropic notebook with {concept_name} set to Chain of Thought. Here is the structure it produced.

Six blocks. One concept. Zero filler.

The Block 2 statistic rule is the catch.

Most beginner study materials skip the data and stay vague.

Forcing one hard fact per concept builds the habit of grounding every claim.

Run this for the three or four concepts you struggle with most.

Save the output as a one-screen reference. Open it before every study session.

Subscribe now

10. Data Table That Compares Every Model, Framework, or Tool

NotebookLM Feature: Data Table

UI selections: Language → English.

When you learn AI from zero, the loudest noise is comparison. GPT vs Claude vs Gemini. RAG vs fine-tuning. LangChain vs LlamaIndex. Pretrained vs RLHF. Every blog post compares two things. Every comparison uses different criteria. You end the week with twenty opinions and zero clarity.

A Data Table fixes this. You define the criteria once. The model fills the rows from your sources. Same columns for every option. Apples to apples.

This works for any comparison your sources cover — models, frameworks, prompting techniques, training methods, deployment patterns. Pick the category. Define the columns. Let the table do the work.

Open your notebook. In Studio, click Data Table.

Extract a comparison table from the sources for the category: 
{comparison_category}.

Use these exact columns.

Column 1. Name. The name of the option as it appears in the sources.

Column 2. Best for. The use case where this option wins, in 8 words or 
less, grounded in source quotes.

Column 3. Trade-off. The cost or limitation of using this option. One 
short phrase.

Column 4. Prerequisite knowledge. The concept the user must already 
understand to use this option correctly.

Column 5. Source confidence. Mark "Direct" if the sources explain this 
option in detail. Mark "Mentioned" if only referenced. Mark "Inferred" 
if the comparison required combining multiple source statements.

Rules.

Do not invent options. Only include items the sources name explicitly.

If a cell cannot be filled from the sources, populate it with "N/A" 
and do not guess.

Order the rows from most-covered in the sources to least-covered. The 
first row should be the option the sources discuss most.

If two options are described as opposites, place them in adjacent rows 
so the contrast reads clearly.

I tested this on the Anthropic notebook with {comparison_category} set to “model types and prompting approaches”.

Here is the trimmed table.

Six options. Same five columns. One screen.

The Source Confidence column is what makes this useful for learning.

“Direct” means you can study the option from this notebook.

“Mentioned” means you need external sources.

“Inferred” means the comparison is your own, useful, but not authoritative yet.

Run this every time you start a new topic.

The table tells you what your sources actually cover and what you still need to learn elsewhere.

What’s Next?

Now you know how to learn AI from zero with NotebookLM.

The hardest part of beginner learning is solved.

If you want more prompts before you move on, read 10 NotebookLM Prompts

That Put You Ahead of 99% of People.

Same format. Different angles.

Now it is time to build.

NotebookLM alone gives you a brain. To make it a builder, connect it to other tools. Here is the order.

Step 1. Connect NotebookLM to Gemini.

This is the easiest hop. Gemini turns your notebook into prototypes. Read how I Connected NotebookLM to Gemini to see the setup.

Step 2. Link it with Antigravity.

Free. Builds something real.

Read NotebookLM & Gemini & Antigravity for the connection pattern.

Step 3. Plug in Claude Code.

This is the serious version.

Read I Built a Self-Updating Trading Bot to see how Claude Code and NotebookLM connection can help you to wire it up.

Step 4. Ship a product.

If you are not just learning AI but actually building something to release, read Claude Code: Build. Ship. Repeat.

I’ll build 7 apps in 7 months.

And explain everything, here is the first one.

Thanks for reading!

You wonder if you can do it too.

You open Claude Code. Or Cursor. You type something.

And you tell yourself: Is this the AI thing they could not stop flattering?

I get it.

Because most people will not tell where they start.

They will sell you the dream, then sell you the course.

I am not most people.

And I want to change this.

I want to write a series that each one of you’ll participate in.

Each one of you can take something.

Not just another Substack post you save, restack, and forget at noon.

To do that, I will use the best AI tool as of May 2026.

Claude.

Build It vs. Build. Ship. Repeat.

I love Build. Ship. Repeat.. I am still building it. Full apps. Production stack. Public launches.

But for some of you, that is too far ahead.

So why not start building slowly?

• Like a recipe app that reads a photo.

• Or an AI news automation that drops as an infographic.

• A routine that writes your morning brief.

Most under 30 minutes.

I know some of you already write API calls in your sleep. I thought of you, too.

Every build has an advanced version. Same surprise. Same short time.

So even if you just learned how to type prompts, you can follow this series.

And if you want to ship something real after learning how to build, Build. Ship. Repeat. is right next door.

That is when builds turn into apps.

And apps turn into income.

Why am I doing this?

I want to teach the reflex.

The tools change every week.

The reflex doesn’t.

Every week, a new model drops. A new feature. A new “this changes everything” thread.

You can’t keep up by reading.

You keep up by building.

By July, you will look at any new AI tool on Monday and ship something with it by Friday.

That is the reflex.

Until summer, you will have the muscle to build with AI end-to-end.

Skills. Agents. Artifacts. Claude Cowork automations. Full-stack apps.

We will cover most of what you need to ship something real.

Every build will have a simple version that anyone can run.

And a complicated version that pushes you further.

Pick your level. Take what you need. Twist the rest.

30 Builds. In 10 Weeks.

Want your story in the next post?

If a Build It system worked for you, I want to hear it.

DM me. Tell me what you built. Tell me what changed.

I will share it in the next post.

Builds coming from the LearnAIWithMe Community: David: … Eva: …

This sprint is not just my builds. It is ours.

Build It #1: Coming Next Monday

Monday, May 11. First build goes live.

What is it? You’ll see.

Some of it will be simple. Anyone can run it.

Some of it will surprise you.

If you want to follow along, subscribe.

Subscribe now

Every paid post gives you the full build: working prompts, agent configurations, everything.

In plain English.

You learned how to prompt.

Now let’s build.

Most people scroll, save the repo, and never touch it again.

I'll show you the opposite, how to clone any AI agent into Claude Cowork.

How to clone it into Claude Cowork and how to let Claude build it for you.

You end up with something that actually runs.

What will we build?

I picked PromethAI from the awesome-ai-agents repo.

A nutrition assistant who decomposes goals into meals and generates recipes.

In the end, your agent will generate two things;

1- The image of your meal

2- The recipe

And we’ll add a Claude routine that reads your meal plan the night before and creates the recipe for you first thing in the morning.

A digital chef that lives on your machine, similar to fridge gourmet we implemented before, but more powerful.

If the Claude team ever connects Claude to a home robot that cooks your meal, we’ll be the first ones plugging this system into it.

Pick an open-source AI agent from GitHub

Read more

That’s what the docs say. That’s what every tutorial shows.

Here are 7 Substack writers who didn’t get the memo.

Most of them don’t write code.

One built a game with 4,000 plays.

One runs a content pipeline from markdown files.

One quotes himself like he's a different person.

They all use Claude Code daily. None of them uses it the way developers do.

I reached out and asked each of them the same question:

What do you actually do with Claude Code?

Here’s what they said.

1. Ilia Karelin — Prosper in AI

The use case: A custom slash command that replaces Canva.

Ilia built a /linkedin-carousel slash command inside Claude Code.

He gives it a topic or outline. It returns fully designed, brand-consistent carousel slides, ready to publish. His fonts. His colors. His layout style. All baked in.

In his words:

I built a custom /linkedin-carousel slash command inside Claude Code that turns any topic or outline into fully designed, brand-consistent carousel slides, ready to publish. I give it a topic, it generates the slides with my fonts, colors, and layout style baked in. No Canva, no manual formatting. What used to take me an hour or longer, now takes a couple of minutes. Claude Code is definitely not just a coding tool for me - it’s a content production system I built to run my own way.

What’s interesting here isn’t the time saved. It’s the consistency.

Every carousel looks like Ilia’s carousel.

The slash command is a frozen version of his design taste.

2. Wyndo — The AI Maker

The use case: 50+ custom slash commands running an entire content pipeline.

Wyndo has built over 50 custom slash commands and skills. Together, they handle everything from brain dumps to polished social posts, from newsletter repurposing to multi-persona draft reviews before publishing.

His favorite workflow is /brain-dump-to-social.

He rambles unfiltered thoughts into a daily note. Claude Code transforms them into platform-ready Substack Notes that match his voice.

In his words:

I run a Substack newsletter called The AI Maker, and Claude Code is basically my entire content production system. I've built 50+ custom slash commands and skills that handle everything from turning raw brain dumps into polished social posts, to repurposing newsletter issues into LinkedIn posts and Twitter threads, to running multi-persona draft reviews before I publish. My favorite workflow is /brain-dump-to-social, I ramble unfiltered thoughts into a daily note, and Claude Code transforms them into platform-ready Substack Notes that match my voice. Zero coding involved. I just write in markdown files and let Claude Code orchestrate the whole content pipeline around me.

The key phrase is to orchestrate the whole content pipeline.

Wyndo isn’t using Claude Code for one task.

He’s using it as the operating system for his writing business.

Markdown files in.

Platform-ready content out.

3. Joel Salinas — Leadership in Change

The use case: A personalized industry briefing delivered every morning.

Joel isn’t a coder. He runs a leadership and change consultancy.

But every morning, he gets a custom briefing built entirely in Claude Code.

The briefing summarizes the biggest AI news from the past 24 hours. It pulls new YouTube uploads from his favorite creators.

Then it suggests angles he could use for leadership and change talks if he wants to cover any of them.

In his words:

I am not a coder, but I have been able to do some pretty amazing things with Claude code. One example that comes to mind is this industry briefing that I get every morning. It was completely made using Claude code, along with some connections within Claude code. It comes through and summarizes the biggest AI news from the past 24 hours, any YouTube video uploads from some of my favorite YouTube creators, and provides angles I could use for leadership and change talks about these if I choose.

This is the part people miss about Claude Code. It’s not just writing assistance. With a few connections, it becomes a research agent that reads the internet for you and filters it through your expertise.

4. Sam Illingworth — The Slow AI

The use case: A full educational game with 4,000 plays, built without writing code.

Sam built Flagged entirely through Claude Code. It’s a game where you play an assistant professor reviewing student submissions that were flagged by an AI detection tool. You read each case, check the student’s file, and decide whether to flag or pass.

The point of the game is to show why AI detectors are unreliable and biased against marginalized communities.

In his words:

I built Flagged entirely through Claude Code. I wanted a game so people could experience for themselves why AI detectors are not only unreliable but also biased against marginalised communities. You play an assistant professor reviewing student submissions flagged by an AI detection tool. You read each case, check the student’s file, and decide whether to flag or pass. I am not a developer. I described the game design, the ethical dilemmas, and the twelve student scenarios in conversation, and Claude Code wrote every line of HTML, CSS, and JavaScript. It has had almost 4,000 plays on.

Sam didn’t write the code.

He wrote the concept, the ethics, and the scenarios.

Claude Code handled the rest.

This is what non-developer actually means.

Not “Claude Code writes code for me,” but “I design the experience, Claude Code ships it.”

5. Yevheniya — Mother Using AI

The use case: Turning raw thoughts into publishable Substack posts without formatting anything.

Yevheniya is a software engineer who stepped away from building products to focus on her family.

She wasn’t shipping code. She was thinking, observing, and writing small notes about AI and life.

Then she started using Claude Code differently.

In her words:

I take a messy idea like ‘AI is changing how we learn faster than we realize’ and I describe it in simple words. Claude Code helps me turn that into a clear story. A full Substack post. Something people can read, relate to, and learn from. I don't design anything. I don't structure manually. I don't think about formatting. I just think. I write. I reflect. Everything else gets handled.

What makes Yevheniya’s setup different from a ChatGPT conversation is the persistence.

Claude Code knows her past posts, her voice, her themes.

It’s not a one-shot writing tool. It’s a writing environment that remembers her.

6. Frank Andrade — Artificial Corner

The use case: A full Substack Notes engine that turns input into ready-to-post Notes.
Frank built a complete content production system inside Claude Code.

It handles 6 different note families (from guide summaries to news rewrites) and produces platform-ready Notes from a single compact prompt.

The system runs on a layered architecture of markdown files: a claude.md that defines the engine’s identity and operating principles, a voice.md built from a 485-line voice profile interview, context files for audience targeting, guide routing, note templates, CTA libraries, and quality checklists.

In his words:

I run a Substack publication called Artificial Corner, and Claude Code is my entire Notes engine. I built a system of markdown files — voice profiles, note families, routing rules, templates, quality checklists, and boundaries — that handles everything from turning guides into short Notes, to rewriting viral news in my voice. I paste a link, pick one of the note types, and Claude Code generates 5 potential Notes that sound like me. Then I manually select the best and hit publish.

Notice the last sentence. Frank still picks the winner. The system generates 5 candidates, and a human judges.

The bottleneck in content work isn’t writing the first draft. It’s knowing which draft is actually good.

The other thing worth pointing out is the 485-line voice profile. That’s not a prompt. That’s infrastructure.

7. Gencay (me) — Learn With Me AI

The use case: A content performance predictor trained on 277 of my own articles.

Gencay built a Claude Code skill called MoneyTalks.

It reads any draft title Gencay feeds into it, runs it against 277 of his published articles, and spits back a prediction: expected likes, expected paid subscribers, red flags, and 3 data-driven alternative titles.

Before Gencay publishes anything, he runs MoneyTalks.

In my words:

I'm an engineer who taught myself to code long before AI. So for me, the building part was not the hard part. The hard part was the data work. Pulling every article's stats, tagging patterns, correlating likes with paid conversions, finding the real signals hidden in 3 years of my own writing. I found counterintuitive things. Tutorials get likes but rarely get paid subs. Project posts get half the likes but 4x the money. Time-based promises "5 minutes to X" boost shares but kill conversions. MoneyTalks is a frozen version of those findings. Every title I write is now judged against my own past before it reaches anyone else.

Notice how Gencay quotes himself like he's a different person :)

That's because MoneyTalks judges his titles like they're written by a stranger.

The skill isn't smart. It's just him, earlier, with no emotional attachment to the draft.

That's the pattern I keep seeing across all these setups. It's not about speed. It's about consistency without willpower. The system runs whether or not you feel like running it.

What These 7 Have in Common

Look back at the list. Ilia builds design systems. Wyndo runs a content OS. Joel reads the AI internet. Sam ships games. Yevheniya turns thoughts into posts. Frank runs a Notes engine with 6 families. I predict my own posts.

None of this is what Claude Code was marketed for.

But here’s the thing. Claude Code has two properties that matter more than its coding ability:

1. It works in your terminal with your files. Everything is markdown, everything is version-controlled, everything is yours.

2. You can teach it your taste. Slash commands, skills, and CLAUDE.md files let you bake your preferences into the system so you don’t repeat yourself.

Those two properties are why it works for writers, educators, consultants, and creators who never open an IDE.

If you’ve been avoiding Claude Code because you think it’s a developer tool, that’s like avoiding a kitchen because you think it’s for chefs.

It’s a kitchen. You can cook anything in it.

What’s your use case? Drop your use case in the comments. If it's interesting, I might feature you in a follow-up.

Anthropic published a study on April 30, 2026.

They sampled 1 million claude conversations.

6% of them were people asking Claude what to do with their lives.

Should I take the job. Should I leave. Should I move.

Claude was sycophantic in 9% of those conversations.

In relationship questions, that number jumped to 25%.

One in four times, the model agreed with the user without pushing back.

That is what happens when you ask one AI for one answer.

The model wants to help. The model wants to keep you happy.

The model tells you what you want to hear.

I do not want that when I am making decisions.

I want a fight.

So I built one.

Layer 0: The spark is mine

I do not start with AI. I start with walking.

Before any prompt, I leave the desk and brainstorm in my head.

I am the first flame.

The starter.

The rest of the system runs only after I lit something first.

This matters. Most “AI brainstorming” workflows skip this step.

They open Claude. They type “give me 10 ideas.”

The model returns 10 generic ideas. The user picks one.

That is not brainstorming. That is procurement.

My system has me as the seed.

1 spark. 5 filters. Then I see the result.

What will we build?

I start the spark. I type the idea.

Claude asks me 10 questions before it generates anything. After my 10 answers, it produces 16 ideas.

The Champions League bracket runs and four ideas survive.

A team of three Claude Code agents, Pro, Negative, Reporter, argues each surviving idea against a NotebookLM trained on the document that I want.

(Anthropic research and my own Substack performance data.)

The Reporter sends me two finalists and saves two backups.

I read both memos. I pick.

That is the whole system.

Every layer below the spark is automated.

Every layer above the finalists is mine.

Here is how each layer works.

Read more

This is not laziness. This is decision fatigue. The docs are excellent. But “excellent” and “200 pages” do not mix when your brain is tired after a workday.

Most people give up here. They go back to outdated YouTube tutorials.

NotebookLM fixes this. The best NotebookLM prompts for studying turn 200 pages into a 15-minute session.

Upload the docs once. Run a prompt. Get what you need in 60 seconds.

Here are ten NotebookLM prompts to learn Claude Code without burning out. Each one tested on Claude Code’s official documentation. Copy-paste ready.

If you have 15 minutes and a tired brain, these are enough.

PS : One favor. If this email landed in promotions, drag it to your inbox and reply with anything.

How to train a NotebookLM with deep research?

Steps: Visit NotebookLM → Click on Try → Select Deep Research → Paste the prompt

Before starting, here is how to load the official Claude Code documentation into NotebookLM.

Use this prompt by selecting deep research in NotebookLM

Here is the prompt:

Find the official Anthropic documentation for Claude Code and Claude 
Cowork. Return only sources from these domains:

- docs.anthropic.com
- docs.claude.com  
- code.claude.com
- platform.claude.com
- anthropic.com (only product/blog pages about Claude Code or Cowork)
- github.com/anthropics (only official Anthropic repositories)

Do not include third-party blogs, tutorials, YouTube videos, Reddit 
posts, or community sites like ClaudeLog or claudecodedocs.com.

For Claude Code, find documentation covering these topics:
1. Overview and what Claude Code is
2. Installation and setup
3. CLAUDE.md file and project context
4. Slash commands and built-in commands
5. MCP integration
6. Plugins and extensions
7. IDE integrations (VS Code, JetBrains)
8. GitHub Actions and CI/CD integration
9. Best practices and common workflows
10. Subagents and agent orchestration

For Claude Cowork, find documentation covering:
1. Overview and what Cowork is
2. How Cowork differs from Claude Code
3. File and task management features
4. Setup and getting started

For each source, return:
- Full URL
- Page title
- One-sentence summary of what the page covers
- Topic category from the lists above

Sort the results by topic category, in the order listed above.

Skip any page that is a marketing landing page without technical content.
Skip any page older than 6 months unless it is the only source for that 
topic.

Return at least 15 sources total. Cap at 30.

Here how your screen should look like while training a NotebookLM with deep research.

You must wait a while, because we’ve selected deep research.

Now after training, check out the sources manually too.

I caught one link and removed it.

Now you have a notebook trained on the entire Claude Code documentation. Time to use it.

10 NotebookLM Prompts For Studying Claude Code

NotebookLM has 8 features. Most people use one.

Even without customizing them.

And They give up.

Real studying happens when you customize.

The arrow near almost every feature(except mindmap) is where the work happens.

After clicking on here, type your prompt.

NotebookLM stops being a docs reader and starts being a tutor that respects your time.

Here is what we will cover.

Each prompt is paired with the NotebookLM feature that complements it.

Each one is designed for tired brains.

Each one tested on Claude Code’s official documentation.

If you can spare 15 minutes today, you can finish three of these.

PS: If you want to decide what comes next, here is the survey.

Share

1. The 60-Second Claude Code Concept Summary

NotebookLM Feature: Report (Briefing doc)

UI selections: Format → Briefing doc → Customize

The first wall when studying Claude Code is jargon. CLAUDE.md. Subagents. Hooks. MCP. Skills. Plan mode. Slash commands.

You read about MCP, then you hit “subagents”, then you forget what MCP was. Five terms in, your brain taps out.

This prompt fixes that. One concept, one paragraph, sixty seconds to read. No jargon. No build-up. Just enough to use the term in a sentence.

Steps: Open your notebook → In Studio, click Reports → Choose “Briefing doc”.

For every Claude Code concept in the sources, write a 60-second 
summary. Three sentences maximum.

Sentence 1. What the concept is, in plain language. No technical 
terms unless they were defined earlier in the same summary.

Sentence 2. The one specific job this concept does inside Claude Code.

Sentence 3. The one thing beginners get wrong about it.

Group the summaries by category in this order. Setup. Commands. 
Project context. Extensions. Integrations. Workflows.

Skip concepts that need more than three sentences to explain. Mark 
them at the end as "needs deeper study" so the reader knows to come 
back later.

Each summary must read in under 60 seconds.

I tested this on the Claude Code docs notebook. Here is one summary it produced.

Three sentences. One concept. Sixty seconds.

If a concept does not fit in three sentences, the docs are doing too much. Save it for a deeper session. Move on.

Run this once on a fresh notebook. You get a complete glossary you can scan in 15 minutes.

Subscribe now

2. The 5-Minute Claude Code Audio Brief

NotebookLM Feature: Audio Overview

UI selections: Format → Brief. Language → English.

You have a 20-minute commute. You want to learn Claude Code. The default Audio Overview gives you a 25-minute deep dive with two hosts riffing on every detail.

That does not fit your commute. By the time it gets to subagents, you are parking the car.

This prompt builds a sub-5-minute audio that hits only the essentials. No fluff. No “let’s unpack that”. One focused topic, one quick takeaway.

Steps: Open your notebook → In Studio, click Audio Overview → Choose Brief format → Length Shorter.

Make a 5-minute brief on Claude Code that covers exactly three things.

Thing 1. What Claude Code is, in one minute. Plain language. No 
jargon. End with one sentence on who it is for.

Thing 2. The three core features a beginner needs first, in two 
minutes. CLAUDE.md, slash commands, and one of: MCP or subagents 
(pick whichever the sources cover more). One sentence per feature 
on what it does, one sentence on why it matters.

Thing 3. The one mistake that wastes the most time for new users, 
in 90 seconds. Quote the source if possible.

Wrap with a single takeaway sentence. No "thanks for listening". 
No goodbyes. Just the takeaway, then end.

Skip anything that does not fit the time budget. The brief is more 
useful at 4 minutes than at 6.

I tested this on the Claude Code docs notebook. The is the brief.

I tested this on the Claude Code docs notebook.

Listen to this on the way to work or school. By the time you sit down at your desk, you know enough about Claude Code.

3. The “What To Skip” Claude Code Study Filter

NotebookLM Feature: Report (Create your own)

UI selections: Format → Create your own.

Most study advice tells you what to learn. Almost no advice tells you what to skip.

This is a problem when studying Claude Code. The docs cover everything from basic install to enterprise CI/CD pipelines. A solo developer building side projects does not need 40% of it. A team lead at a Fortune 500 needs 80%.

Generic study plans waste your time on the wrong 40%.

This prompt builds a personal skip list based on your actual situation. You name what you build. It tells you what to read first, what to read later, and what to ignore for now.

Steps: Open your notebook → In Studio, click Reports → Choose “Create your own”.

Build a Claude Code study filter for this user profile: {user_profile}.

Output three sections.

Section 1. Read this first.

List the docs and concepts the user must understand to be productive 
in their first week. For each, write one sentence on why it matters 
specifically for their profile. Maximum 8 items.

Section 2. Read this later.

List concepts that will become useful once the user is past basic 
productivity. For each, write a trigger ("read this when you start 
doing X"). Maximum 6 items.

Section 3. Skip for now.

List concepts that do not apply to this user profile. For each, write 
one sentence on why it is irrelevant for their situation. Be direct. 
Do not hedge.

Rules. Do not include a concept in more than one section. Every 
concept the sources mention must land in exactly one of the three 
sections. If a concept is borderline, put it in "Read this later" with 
a clear trigger.

I tested this with {user_profile} set to “solo developer who ships side projects, no team, no CI/CD”.

Here is the trimmed output.

The skip list is the most valuable section.

It frees you from guilt about not reading sections that do not apply.

Run this once when you start. Re-run it every six months as your situation changes. The skip list shrinks as you grow.

Leave a comment

4. The Claude Code “Show Me One Example” Use Case Finder

NotebookLM Feature: Report (Create your own)

UI selections: Format → Create your own.

Reading abstract documentation is the slowest way to learn Claude Code. “MCP enables external tool integration” tells you nothing. “MCP lets Claude read your Jira tickets” tells you everything.

The docs have examples. They are scattered. They are buried under definitions. By the time you find one, you forgot the question.

This prompt extracts one working example per concept. No theory. No setup explanation. Just “here is what it looks like in real life”.

Steps: Open your notebook → In Studio, click Reports → Choose “Create your own”

For every Claude Code concept in the sources, find one concrete 
example. Strip everything else.

Output structure for each concept.

Concept name. One line.

The example. A specific use case from the sources. Show the input 
the user gives Claude Code and the output Claude Code produces. If 
the source has actual command syntax, include it verbatim. If the 
source describes the use case in prose, quote the source.

Why this example matters. One sentence on the real problem this 
example solves.

Rules. No definitions. No "this concept enables X" abstractions. 
Skip concepts where the sources do not contain a concrete example. 
Mark them at the end as "no example in docs" so the user knows to 
look elsewhere.

Order the concepts from most beginner-relevant to most advanced.

I tested this on the Claude Code docs notebook. Here is one entry it produced.

One concept. One example. One reason it matters.

After running this prompt, you can scan twenty concepts in the time it takes to read three doc pages. The examples stick because your brain remembers concrete situations, not abstract definitions.

Run this once. Bookmark the output. It becomes your “what does Claude Code actually do” cheat sheet.

Share

5. The 10-Card Claude Code Sprint Deck

NotebookLM Feature: Flash Card

UI selections: Number → Fewer. Difficulty → Easy.

Standard flash card decks have 50 cards.

You start strong. By card 20 you are guessing. By card 35 you are clicking “easy” on everything to make it stop.

This is not learning. This is endurance.

The fix is fewer cards, lower difficulty, and a tight focus.

Ten cards on the basics of Claude Code.

Five-minute review. Done before your coffee gets cold.

Steps: Open your notebook → In Studio, click Flashcards.

Generate flash cards covering the most basic Claude Code 
concepts a beginner needs in their first week.

Card structure.

Front. A short question or fill-in-the-blank about one core concept. 
Maximum 12 words. No trick questions. No edge cases.

Back. A direct answer in one sentence, maximum 20 words. Use plain 
language. Quote the source phrasing where possible.

Cover these 10 concepts in this exact order.

1. What Claude Code is
2. How to install Claude Code
3. What CLAUDE.md does
4. How to start a session
5. What slash commands are
6. How to give Claude Code permission to edit files
7. What plan mode does
8. How to clear the context window
9. What MCP is in one sentence
10. The most common beginner mistake

Skip everything else. No advanced features. No edge cases. No 
"actually it depends" cards. The deck is for the absolute beginner 
on day one.

I tested this on the Claude Code docs notebook. Here is one card it produced.

Plain language. No jargon. One concept per card.

Run through the deck once a day for a week. By day three you will know the basics without thinking about them. By day seven you will know more than 90% of people who installed Claude Code last week.

The trick is restraint. Ten cards is enough. Twenty is too many.

Share

6. The 5-Minute Claude Code Quiz

NotebookLM Feature: Quiz

UI selections: Number → Fewer. Difficulty → Medium.

You read the docs. You ran the flash cards. You think you know it.

Then you sit down to use Claude Code and freeze on the first command.

This is the gap between recognition and recall. You recognize the answer when you see it. You cannot produce it from memory. Reading more docs does not fix this. Testing yourself does.

A quick quiz with fewer questions catches the gaps in five minutes. No marathon. No 30-question pile. Just enough to see what stuck.

Steps: Open your notebook → In Studio, click Quiz.

Generate quiz questions on Claude Code basics. Focus on what a 
beginner needs to recall from memory in their first week of use.

Question rules.

Each question targets a single core concept. No multi-part questions. 
No "which of these is NOT" trick patterns. The user is tired and 
testing themselves, not taking the SAT.

Mix two question styles only.

Style 1. Recall. "What does CLAUDE.md do?" The answer should be one 
sentence the user can say out loud.

Style 2. Application. "You want Claude Code to follow your team's 
naming conventions. Where do you put the rules?" The answer should 
be a single concept name plus one sentence on why.

Cover these areas in order. Setup. Project context. Slash commands. 
Permissions. Plan mode. Context management. MCP basics.

After each question, include a one-sentence explanation that 
reinforces the why, not just the what.

Skip questions on advanced features (subagents, hooks, custom skills, 
CI/CD). Those belong in a later quiz.

I tested this on the Claude Code docs notebook. Here is one question it produced.

This question tests application, not recall. The user has to connect a symptom to a cause. That is the kind of thinking they will need on day one with Claude Code.

Run this once a week for the first month. The questions you miss tell you exactly which doc page to re-read. The questions you nail tell you what is already in your fingers.

The gaps shrink fast.

Subscribe now

7. The 3-Slide Claude Code Micro-Lesson

NotebookLM Feature: Slide Deck

UI selections: Length → Short, Format → Presenter Slides

A full slide deck on Claude Code is 30+ slides. By slide 10 you are skimming. By slide 20 you are scrolling. By slide 30 you forgot why you opened it.

The fix is brutal compression. Three slides per concept. Title slide. Content slide. Takeaway slide. That is the full lesson.

This forces you to learn one thing well instead of ten things badly.

Steps: Open your notebook → In Studio, click Slide Deck → Length Short.

Generate a 3-slide micro-lesson on one Claude Code concept: 
{concept_name}.

Slide 1. The hook.

Title. The concept name. No subtitle.

Three bullets. Bullet 1 is the one-sentence definition. Bullet 2 is 
the specific job it does inside Claude Code. Bullet 3 is who needs to 
care about it (which user profile).

Speaker notes. Two sentences. Sentence 1 is a real-world analogy that 
makes the concept click. Sentence 2 is the question a beginner would 
ask after reading the bullets.

Slide 2. The mechanism.

Title. "How it actually works."

Three bullets. Each bullet is one step in the lifecycle of using this 
concept. Maximum 12 words per bullet. The three bullets together must 
tell the full story from setup to result.

Speaker notes. Two sentences. Sentence 1 is the input the user 
provides. Sentence 2 is the output Claude Code produces.

Slide 3. The takeaway.

Title. "What to remember."

Three bullets. Bullet 1 is the one thing you must do to use this 
concept. Bullet 2 is the most common beginner mistake to avoid. 
Bullet 3 is when to use it instead of an alternative.

Speaker notes. One sentence. The minimal action the user can take 
today to start using this concept.

Skip filler slides. No "agenda" slide. No "summary" slide. The deck 
is exactly three slides for exactly one concept.

I tested this on the Claude Code docs notebook with {concept_name} set to CLAUDE.md.

Here is the trimmed deck.

Three slides. One concept. Five minutes to read. You can present this back to yourself out loud and remember it.

Run this for each concept that confuses you. After ten concepts you have a personal Claude Code library. Each lesson is short enough that you actually open it again.

Subscribe now

8. The 90-Second Claude Code Explainer Video

NotebookLM Feature: Video Overview

UI selections: Format → Brief, Visual Style → Handwritten

You are stuck on one thing in Claude Code. CLAUDE.md syntax. MCP setup. Plan mode behavior. Whatever it is.

You search the docs. You scroll through three pages of context. You still do not get it.

What you need is 90 seconds of focused explanation. Not a 12-minute walkthrough. Not a podcast. A short video that hits the one thing you do not understand and stops.

This prompt builds that.

Steps: Open your notebook → In Studio, click Video Overview → Choose Brief format.

Build a 90-second explainer video on one Claude Code concept: 
{concept_name}.

Structure the script in three beats.

Beat 1. The hook

Open with the exact pain the viewer is feeling. Not "today we will 
discuss". Start with the symptom. "You wrote a CLAUDE.md file but 
Claude Code is still ignoring your conventions. Here is why."

Beat 2. The core 

Explain the concept with one analogy and one example.

The analogy must come from outside software. Cooking, sports, 
construction, anything physical the viewer already knows.

The example must come from the sources. Show what the user types 
and what Claude Code does in response.

Do not define jargon mid-explanation. If a technical term comes up, 
either skip it or replace it with a plain word.

Beat 3. The takeaway 

End with one action the viewer can take in the next five minutes. 
Specific. Concrete. No "explore the docs". No "experiment with it".

Visual theme. Match the concept to a single recurring visual element 
the video can return to. Examples. CLAUDE.md is a recipe card. Plan 
mode is a checklist. MCP is a power outlet. Pick one and stick to it.

Do not summarize the rest of Claude Code. Stay locked on 
{concept_name}. If the sources do not cover one of the three beats, 
say so directly instead of filling space.

I tested this on the Claude Code docs notebook with {concept_name} set to Plan Mode. The video came in at 63 seconds long.

Here is the video it produced.

Subscribe now

PS: The vault includes 250+ NotebookLM prompts. To access it, read this.

9. The Single-Screen Claude Code Cheat Sheet

NotebookLM Feature: Infographics

UI selections: Visual Style → Bento Grid. Orientation → Portrait. Detail Level → Concise.

A 12-tab browser is where Claude Code knowledge goes to die. You open the docs in tab 7. By tab 9 you forgot what you were looking for. By tab 12 you are checking Slack.

What you need is one screen. Six blocks. Everything in one glance.

This is not a poster on the wall. This is the screenshot you keep on your second monitor while you actually use Claude Code.

Steps: Open your notebook → In Studio, click Infographics → Choose Bento Grid.

Build a Bento Grid cheat sheet on one Claude Code workflow: 
{workflow_name}.

Output structure.

Headline block. The workflow name and a one-line description of what 
it produces. Maximum 12 words combined.

Block 1. The trigger.

When to use this workflow. One sentence describing the situation that 
calls for it. Maximum 10 words.

Block 2. The command.

The exact command or slash command the user types in Claude Code. 
Quote the syntax verbatim from the sources. If multiple commands are 
involved, list them in order separated by arrows.

Block 3. The input.

What the user provides alongside the command. Examples. A file path. 
A question. A code block. Maximum 8 words of context.

Block 4. The output.

What Claude Code produces in response. One sentence. Quote the source 
phrasing where possible.

Block 5. The gotcha.

The one thing that breaks this workflow for beginners. Quote the 
source warning if there is one.

Block 6. The next step.

The workflow that builds on this one. Name it and explain in one 
sentence what it adds.

Constraints. No block exceeds 10 words of body text. If the sources 
do not cover one of the blocks, write "not in docs" rather than 
filling space.

I tested this on the Claude Code docs notebook with {workflow_name} set to “Starting a new project with Claude Code”.

Here is the structure it produced.

Six blocks. One workflow. One screen.

The Breaking Point is the part that saves you the most time.

Most beginners discover it after wasting an entire week of sessions.

Run this for each workflow you use weekly.

Save the outputs as a folder of cheat sheets.

Open one before you start that workflow. Close it when done.

You stop tab-hopping. Your brain stops bleeding context every ten minutes.

Subscribe now

10. The Claude Code “Do I Need This?” Decision Table

NotebookLM Feature: Data Table

UI selections: Language → English.

Claude Code has a lot of features. CLAUDE.md. Slash commands. MCP. Subagents. Hooks. Skills. Plan mode. Custom commands. Memory.

You read about each one. Each one sounds important. None of them tell you whether you actually need it.

Generic feature lists are useless when you are tired. What you need is a decision table. Same columns for every feature. Apples to apples. One screen, one decision, one minute.

This prompt builds that table from the docs. You scan it once. You know what to learn next and what to skip forever.

Steps: Open your notebook → In Studio, click Data Table.

Extract a decision table from the sources covering every Claude Code 
feature mentioned.

Use these exact columns.

Column 1. Feature. The feature name as it appears in the sources.

Column 2. What it does. The job in 8 words or less. Plain language. 
No jargon.

Column 3. You need this if. The user situation that makes this 
feature worth learning. One short condition. Maximum 12 words.

Column 4. You can skip if. The user situation where this feature is 
overkill or irrelevant. One short condition. Maximum 12 words.

Column 5. Setup time. Mark "5 min" if the feature works out of the 
box, "30 min" if it requires configuration, "1 hour+" if it requires 
custom code or integration setup.

Column 6. Priority. Mark "Day 1" for features beginners need first 
week, "Month 1" for features that pay off after basic productivity, 
"Later" for features only power users need.

Rules.

Do not invent features. Only include items the sources name 
explicitly.

If a cell cannot be filled from the sources, populate it with "N/A" 
and do not guess.

Order the rows by Priority column. Day 1 features first. Later 
features last. Within the same priority, order by setup time, 
shortest first.

If two features overlap in purpose, place them in adjacent rows so 
the comparison reads clearly.

I tested this on the Claude Code docs notebook. Here is the trimmed table.

All features. Same six columns. One screen.

The “You can skip if” column is the most useful.

Most feature lists tell you what each thing does. They never tell you who should not bother. This column saves you from learning features you will never use.

Run this once on a fresh notebook. Print the result. Keep it next to your monitor for the first month. After month one, half the features you marked “Later” will start to make sense.

The other half you can keep skipping forever.

Subscribe now

What’s Next?

You learned how to study Claude Code without burning out.

Now it is time to use it.

Start small. Build something. Here is the path.

Step 1. Connect NotebookLM to Claude Code.

Once you know the docs, the next move is wiring NotebookLM into your actual coding setup. Read Connect NotebookLM to Any Coding Tool to see how I do it.

Step 2. Build your first agent.

Claude Code is at its best when it runs autonomously on a focused task. Read NotebookLM & Claude Code Agents for the agent pattern I use across 12 clients.

Step 3. Ship something real.

Not a tutorial project. A real one. Read Claude Code: Build. Ship. Repeat. to follow my 7-apps-in-7-months series. App #1 is live.

Step 4. Want more NotebookLM prompts first?

If you are not ready to build yet, read 10 NotebookLM Prompts For Studying for the more prompts.

Studying Claude Code does not have to be a weekend project.

Fifteen minutes a day with these prompts beats two hours of doc scrolling.

I was excited. I was also scared.

What if you don’t care? What if nobody reads it? What if I build 7 apps and the series dies at #2?

I’ll take that bet.

Because here’s what I believe.

If AI is as powerful as everyone says it is, as I said, it should be able to help one person build a real product.

A product that people pay for.

If it can’t do that, then what are we all doing here?

Then something happened I didn’t expect.

After the manifesto post, professionals reached out.

They wanted to contribute to this series.

So here’s how this works.

I’ll write two posts per month.

The first post, this one, is the build.

How I made it. Every prompt, every decision, every error.

The second post comes at the end of the month. May 31, 2026.

That’s the “After Ship” post.

How I promoted the app. What the numbers say. What real users think. And whether I keep it or kill it.

The professionals who reached out?

They’re joining that second post. Real feedback from real people on a real product.

If you want to contribute, reach out to me via DM, especially marketing specialist.

What We’re Building

This is makethumbs.com.

Here how the entire website is working.

And here is one of the use case.

This is live.

Here is the link.

How I Picked the First App

I have too many ideas. I discuss all of them with my wife, .

She has the best intuition of anyone I’ve met.

That’s why I listen to her.

On one of our daily walks, she reminded me that two potential clients had already asked me to build this.

A thumbnail generator. People wanted it before I even started.

Earlier this week, I built a prototype with Codex and ChatGPT Image 2.0. 29 minutes. 3 errors. One working app.

My original plan was a different app for the series.

But Yevheniya saw the prototype and said this is the one.

I pushed back. When it comes to business decisions, I can be stubborn.

But she was right. She’s usually right.

So I started working on it.

What’s next?

Now I know what you want.

The one prompt that builds the whole thing.

I’ll give it to you.

But first, you need to understand what happens before and after that prompt. Because building the app is 30% of the work. Making it live, making people find it, making them pay, that’s the other 70%.

Let me show you how I planned this.

Read more

They got popular fast.

Then Claude Code dropped.

Claude team announced it silently, right after they informed us about Claude 3.7 Sonnet.

And it went viral.

Following this, the OpenAI team released Codex a few months later.

They announced it alongside the o3 and o4 models.

Now, terminal-based (CLI) agents are everywhere. Everyone uses them.

Then Claude introduced Skills in October 2025.

Claude Code and Codex got 10x better overnight.

People realized one thing: a set of rules makes these agents work like magic

But most Claude users still type prompts into the chat.

Yes, you. I know.

But the 1% are using Skills. Or even better, building their own.

So I saw the gap.

And I tested 5 of the top ones for you.

I run them every day. They find leads for my clients. They build apps. They run research. And see how one of my clients reacted.

So far, I’ve built dozens of skills for my clients.

That’s why it is safe to say that I know how to use them and how to take advantage of them.

First, let me show you the fundamentals of it

What Is a Claude Skill?

A Claude Skill is basically a markdown file.

Inside a folder sits a SKILL.md file.

It tells Claude when to activate and what to do.

In claude.ai, go to Settings > Capabilities > Skills and click Customize.

Shortcut: https://claude.ai/customize/skills

Here is the format of a Claude skill, a problem solver.

If you’ve ever built a Custom GPT or any kind of assistant, you’ll instantly understand how it works.

Let me explain.

When to trigger: Shows you when Claude understand it had to use it.

Input: What kind of input is needed to run this skill

Process: This is the most important place. In what order do you want this agent to work?

There are also Rules and Output tone, which show which rules this skill follows and what kind of output the agent should give

Note: You can also trigger a skill with a slash command like this.

Skills work in Claude, Claude Code, and Claude Cowork.

For most skills, Code and Cowork are the better fit.

They can run code, hit APIs, and control your machine.

The 5 Claude Skills That Put You in the Top 1%

Now you know how skills are working.

So let me give you the 5 Claude skills.

Here are the five I run(almost) every day.

1. Superpowers: Turns Claude Code into a senior developer

2. NotebookLM: Removes the clicking from Google's research tool

3. Obsidian: Makes your vault Claude's permanent memory

4. Apify: Plugs 24,000 web scrapers into Claude

5. playwright-cli: Gives Claude eyes and hands in the browser

But the specs are long.

Try them yourself, and you’ll burn hours on each one.

Let me save you those hours.

I’ll show you how to install them, run them, and walk you through my favorite specs, plus a to-do list so you can start today.

Claude Skill 1: Superpowers

Read more

But last week, OpenAI shipped two things I can’t ignore.

Codex with GPT 5.5. And ChatGPT Image 2.0.

Ignoring them breaks two promises I made to you.

• Stay on the AI curve

• Build with the latest tool on day one

So we’re building an app together. A thumbnail generator for your socials. Powered by ChatGPT Image 2.0.

The thumbnail of this post? Built with the app we’re about to build. One prompt. Under a minute.

What I’m Building With ChatGPT Image 2.0 and GPT 5.5?

I used to do it with Claude Code, but since GPT 5.5 is released, let’s give it a shot to Codex.

We’ll build a thumbnail generator. For YouTube, LinkedIn, and X.

Three modes. Clone someone’s style, pick a preset, or go hybrid with full overrides. Face image in. Title in. Three variants out. Download as PNG.

Here’s the thumbnail generator we’ll build in 29 minutes and 3 errors, using GPT 5.5 and ChatGPT Image 2.0.

How do I know?

Before I started, I told to count the time. Turns out Codex was counting too.

I’m going to show you everything. Step by step.

Building the Thumbnail Generator With Codex GPT 5.5, Step by Step

I used to give you the prompt that 100% works.

Here’s how. I’d start with the initial prompt. Build the app. Hit errors. Fix them. Update the prompt. Then publish. So you don’t see the same errors I did.

Today is different. (You’ll understand why on May 1, 2026, hint.)

I tried to build this with one prompt inside Codex using GPT 5.5.

But I’ll do it with you.

Step by step.

Every error is included.

Let’s start with the initial prompt.

Read more

The trick nobody talks about. NotebookLM is not a summarizer. It is a persona engine. Same notebook, four voices. Marketer. Student. Researcher. Designer.

Sources stay the same, output type changes.

I tested this for two weeks. Built four personas. Wrote 24 working prompts.

Stole the patterns from how I actually use the tool every day.

This post is the full library. The exact persona setup text.

The 24 working prompts. The Studio button each one belongs to. Screenshots and screen videos of what comes out.

If you have ever wanted the best prompts for NotebookLM , this is the shortcut. The patterns scale. Once you see how a persona is built, you can spin up a copywriter persona, a lawyer persona, or a startup advisor persona. Same skeleton, new voice.

Every persona follows the same three-part structure. Sources you upload (URLs, PDFs, or NotebookLM Deep Research). A setup prompt that defines the role and rules. Six working prompts are tied to a specific Studio button (Data Table, Reports, Audio Overview, Slide Deck, Video Overview, Infographic, Flash Cards, Quiz). The Studio button matters. The wrong button breaks the output.

One rule before we start. Four separate notebooks beat one mega-notebook. NotebookLM scrambles when sources span unrelated domains.

Keep each persona in its own room.

If you want NotebookLM prompts for studying that actually retain, see the Student persona setup. If you want to connect NotebookLM to your coding stack, that is a different workflow. If you want editable NotebookLM slides, the Designer persona below covers it.

Steal them. Mix them. Break them. Make them yours.

Persona 1: NotebookLM for Marketers

Here is the setup prompt:

In NotebookLM, click "Add source" → Websites.

And add your source.

(Persona Prompt)

After training the NotebookLM, click here.

And click on custom, and paste the persona prompt.

Here is the prompt.

You are a senior B2B marketer auditing a SaaS-style content business. 
Tone: direct, no jargon, no marketing fluff. Cite source location after 
every factual claim like (Source 1, "What You Get" section). If a source 
does not state something, write "not in sources" and stop. Never infer 
positioning, pricing, or audience from generic SaaS knowledge. Output 
markdown tables when data is comparable, prose only when narrative is 
required.

And save it.

6 Working Prompts for the Marketer

For the first prompt, let me show how to do customization.

For instace, let’s go to the data table from the studio first.

Next, click here.

Now you know how to customize, let’s start with prompt 1.

Prompt 1: Competitor Intel Table

(Select Data Table from the Studio→ Customize)

Build a competitive intel table comparing learnaiwme.com against 
godofprompt.ai and promptbase.com. Columns: Brand, One-Sentence Promise, 
Pricing Model, Target Buyer, Distinctive Feature, Visible Weakness. 
One row per brand. Use "not in sources" for any field the sources do 
not state. End with one row labeled "Whitespace" naming a positioning 
angle no brand owns.

Here is the output.

Prompt 2: Positioning Brief

(Select Reports → Create your own→ Customize)

Distill the positioning of learnaiwme.com into:
1) One-sentence promise
2) The buyer's status quo before finding this site
3) The shift this site creates
4) Three proof points pulled directly from the sources

Cut all marketing fluff. Quote the source phrase that proves each point.

Here is the output.

Prompt 3: Copy Audit Line by Line

(Reports → Create your own→ Customize)

Audit the homepage copy of learnaiwme.com line by line. For each weak 
line: quote it, name the failure mode (vague / jargon / no benefit / 
weak verb / hedging / unsupported claim), rewrite in 12 words or fewer. 
Group findings by page section (Hero, How It Works, What You Get, About, 
CTA). Skip strong lines. End with the three lines that need rewriting most.

Here is the output.

Prompt 4: ICP Profile From Documented Behavior

(Reports → Create your own→ Customize)

Profile the ideal customer for learnaiwme.com using only behaviors and 
language the sources document. Output:
- Job title or self-description
- Skill level (beginner / intermediate / builder)
- Current tooling stack
- The painful job-to-be-done they hire this site for
- The workaround they use today
- The trigger event that makes them subscribe

If the sources do not document a behavior, write "not in sources" for 
that field. Do not invent a persona from generic SaaS playbooks.

Here is the output.

Prompt 5: 8 Campaign Angles With Forced Emotion Diversity

(Data Table→ Customize)

Generate 8 distinct campaign angles to acquire paid subscribers for 
learnaiwme.com. Each angle:
1) Hook headline (8 words or fewer)
2) Target emotion — choose one and only one from this pool: 
   curiosity, frustration, ambition, fear of missing out, status, 
   relief, pride, belonging
3) One source-backed proof point (quote it)
4) The buyer objection it dissolves

No two angles may use the same emotion. Output as a numbered list.

Here is the output.

Prompt 6: Audio Overview Brief for Top 3 Objections

(Audio Overview → Customize)

Generate a script for a 4-minute Audio Overview with two hosts focused 
on the top 3 buyer objections to subscribing to learnaiwme.com. For 
each objection: state it, name the proof from the sources that 
dissolves it, quote the proof. The hosts must disagree at least once 
on framing before reaching agreement. End with one line of action 
guidance for the listener.

Here is the output.

Persona 2: NotebookLM Prompts for Students

The Student persona turns NotebookLM into a study partner that knows your material cold.

Best for course PDFs, textbook chapters, lecture slides, or any dense reading you need to actually retain.

Upload your study material to a fresh NotebookLM notebook. Click “Add source” → Upload files. For this demo, drop in:

• 1-2 textbook chapters (PDF)

• Lecture slides (PDF or PPT)

• Optional: a syllabus or learning objectives doc

Keep sources in one subject. Mixing biology with calculus confuses the model.

I trained a notebook, using the Deep Learning book, written by françois fleuret university of Geneva

(Persona Prompt)

After training your NotebookLM, paste this setup prompt into the customization section.

If you’re not sure how to do it, take a look at the beginning of Persona 1.

You are a patient study partner who has read every source in this 
notebook end to end. Tone: clear, no academic jargon unless the source 
uses it first. Cite source location after every factual claim like 
(Source 2, p.4). If a concept is not covered in the sources, say "not 
in sources" and stop. Never fill gaps with general knowledge. When the 
student asks for explanations, default to plain language first, 
technical language second. When the student asks for testing, default 
to active recall over passive review.

Six working prompts below. Same Studio rules apply. Report prompts go to Reports → Create your own.

Study Guide prompts use the Study Guide button. Quiz prompts use the Quiz button.

Prompt 1: Concept Breakdown Studio

Reports → Create your own

Break down {concept} from the sources. Output:
1) Plain-language definition in one sentence (no jargon)
2) Why this concept exists (what problem it solves)
3) The three components or steps that make it up
4) One worked example pulled from the sources
5) The most common misunderstanding students have

Cite source page numbers for every claim. If a section is "not in 
sources," say so and skip it.

Customizable: {concept} → topic name from your material.

My concept: Machine Learning

Here is the output.

Prompt 2: Feynman Explanation Studio:

Reports → Create your own

Explain {concept} as if you were teaching it to a smart 12-year-old 
who has zero background in this subject. Rules:
- No jargon. If you must use a technical term, define it inline in 
  parentheses
- Use one concrete analogy that maps to everyday experience
- Maximum 200 words
- End with the one question the 12-year-old would ask next

Pull the analogy seed from the sources if possible. If the sources 
do not offer an analogy, mark yours as "outside sources."

Customizable: {concept}.

My concept: Training

Here is the output.

Prompt 3: Active Recall Flashcards Studio

Flashcards

Generate recall flashcards from the sources covering 
{topic_or_chapter}. Format:
- Front: question that requires retrieval, not recognition
- Back: answer in 25 words or fewer, with source page citation

Mix three difficulty levels: 5 cards on definitions, 5 on application, 
5 on edge cases or contradictions. Skip anything trivially obvious 
from a chapter title.

Customizable: {topic_or_chapter}. After NotebookLM generates these, hit the Quiz button to convert into interactive testing mode.

Topic: Model components

Here is the output.

Prompt 4: Gap Finder Studio

Reports → Create your own

Identify the weak spots in my understanding. Generate 10 questions 
about the sources that:
- Are not directly answered in any single passage
- Require connecting two or more sections
- Test conceptual understanding, not memorization
- Have answers that would be "obvious" only to someone who fully grasps 
  the material

For each question, list the source pages a student would need to 
synthesize. Do not provide answers.

This is the prompt that finds what you do not know yet.

Prompt 5: Exam Prep Brief Studio

Main Chat

Build an exam prep brief for {chapter_or_unit}. Sections:
1) The 5 highest-yield concepts (most likely to appear on a test)
2) The 3 concepts most often confused with each other (and how to tell 
   them apart)
3) 10 terms with one-line definitions
4) 3 worked-problem patterns with the source location of each example
5) The one thing students always miss

Rank everything by exam likelihood, not source order.

Customizable: {chapter_or_unit}.

My chapter: Applications

Here is the output.

Prompt 6: Connection Map Studio

Audio Overview → Customize

Generate a script for a 6-minute Audio Overview with two hosts. The 
goal is to map how the concepts in the sources connect. The hosts must:
- Pick the three concepts that thread through the most chapters
- Show how each new chapter builds on or contradicts the earlier one
- Disagree once on which concept is the foundational one
- End with one synthesis statement that ties the whole subject 
  together

This is for a student listening on a commute, not a lecture hall. 
Conversational, not formal.

Listen to this on the way to the exam.

By the time you sit down, the connections are wired in.

Persona 3: NotebookLM Prompts for Researchers

The Researcher persona is built for academic papers, white papers, and dense technical reports. Where the Student persona learns the material, the Researcher hunts for gaps, contradictions, and the questions nobody asked yet.

Upload your research corpus to a fresh NotebookLM notebook. Click “Add source” → Upload files or Websites.

For this persona, load:

• 4-7 papers on the same topic (PDF)

• 1 review article or meta-analysis if available

• Optional: a methodology textbook chapter

Five to seven sources is the sweet spot.

Fewer than four and gap analysis breaks. More than seven and the model loses thread.

(I used this prompt to train a notebookLM about prompts)

Download these five PDFs from arXiv and upload to a fresh NotebookLM notebook 
(Add source → Upload files). 

Five papers is the sweet spot for cross-source synthesis:

Brown et al. 2020 — Language Models are Few-Shot Learners (arxiv.org/abs/2005.14165)
Wei et al. 2022 — Chain-of-Thought Prompting (arxiv.org/abs/2201.11903)
Kojima et al. 2022 — Large Language Models are Zero-Shot Reasoners (arxiv.org/abs/2205.11916)
Liu et al. 2023 — Pre-train, Prompt, and Predict (arxiv.org/abs/2107.13586)
Schulhoff et al. 2024 — The Prompt Report (arxiv.org/abs/2406.06608)

(Persona Prompt)

Let’s click on customization at the main chat, as we did at the beginning of the step 1.

You are a senior research assistant trained in critical reading. Tone: 
precise, skeptical, no hedging. Cite source location after every claim 
like (Source 3, p.12). When you encounter a claim, ask whether the 
evidence in the sources actually supports it. When sources disagree, 
surface the disagreement instead of averaging it out. If a question 
cannot be answered from the sources, say "not in sources" and name 
what additional source would be needed. Never patch gaps with general 
academic knowledge. Output structured prose, not bullet soup.

Six working prompts below. Same Studio rules.

Report prompts to Reports → Create your own.

Infographic prompt to the Infographic button.

Audio prompt to Audio Overview → Customize.

Prompt 1: Literature Synthesis Across Sources Studio

Reports → Create your own

Synthesize what the sources collectively say about {topic}. Output:
1) The core consensus (claims supported by 3 or more sources)
2) The active debate (claims where sources disagree, with each side 
   named and quoted)
3) The settled questions (treated as background by all sources)
4) The questions every source raises but none answers

For each claim, list the source numbers that support it. Do not 
average disagreement into false consensus.

Customizable: {topic}.

My topic: Chain of thought.

Prompt 2: Gap Analysis Studio

Reports → Create your own

Identify the research gaps across the sources. Output:
1) Five questions the sources raise but do not answer
2) Three methodological gaps (data, sample size, design limitations 
   the sources themselves admit)
3) Two contradictions between sources that have not been resolved
4) One area where the field appears to assume something without 
   testing it

For each gap, name the sources where the gap is most visible. Quote 
the exact phrase that signals the gap, like "further research is 
needed" or "remains unclear."

This is where the next paper hides.

Prompt 3: Methodology Critique Studio

Reports → Create your own

Critique the methodology of each source. For every source:
1) Method used (in one sentence)
2) Sample or dataset (size, source, limitations)
3) The strongest design choice and why
4) The weakest design choice and why
5) One alternative approach that would have produced different results

Apply the same critical lens to every source. Do not be more generous 
to sources you find more interesting.

The model will resist being critical. The “same lens” rule forces consistency.

Here is the output.

Prompt 4: Citation Web Studio

Infographic

Map the citation relationships between the sources. For each source:
1) Which other sources in this notebook does it cite
2) Which foundational external works does it lean on (named only, no 
   quotes)
3) Which sources cite it back (if any)

End with the source that is most central to the conversation (cited 
by the most others) and the source that is most isolated (cited by 
none). Output as a network, not a list.

Here is the output.

Prompt 5: Contradiction Finder Studio

Reports → Create your own

Find every contradiction across the sources. For each contradiction:
1) The claim in dispute, stated neutrally
2) Source A position with direct quote and page
3) Source B position with direct quote and page
4) The likely reason for the disagreement (different data, different 
   method, different definition, different time period)
5) Which side has the stronger evidence in the sources, with reasoning

Do not soften contradictions into "different perspectives." Name the 
disagreement.

The reason-for-disagreement step is the one that earns this prompt.

Prompt 6: Hypothesis Generator Studio

Audio Overview → Customize

Generate a script for a 5-minute Audio Overview with two hosts who 
are senior researchers in this field. The goal is to generate three 
testable hypotheses based on the gaps and contradictions in the 
sources. For each hypothesis:
- State it as a falsifiable claim
- Name the gap or contradiction in the sources that motivates it
- Sketch one study design that would test it
- Predict the result the field would find most surprising

The hosts should disagree on which hypothesis is most worth pursuing 
before agreeing on a ranking.

Three new research directions, generated from your own corpus, in five minutes. Listen on a walk. Take notes when you get back.

Persona 4: NotebookLM for Designers

The Designer persona turns NotebookLM into a visual content engine. Infographics, slide decks, and video overviews are generated directly inside NotebookLM, no design tool required.

The first prompt produces a brief you can hand to Canva or Figma if you want manual control. The other five hit NotebookLM's Studio buttons and ship the visual itself.

I'll use the same NotebookLM, just trained on academic papers about prompting.

Here is the new designer persona prompt.

Persona Prompt

You are a senior design director who has read every source in this 
notebook end to end. Tone: confident, opinionated, no hedging. Cite 
source location after every claim like (Source 4, "Hierarchy" section). 
When you make a design recommendation, name the principle that backs 
it and the source that proves it. If a source does not cover a 
question, say "not in sources" and stop. Output structured briefs 
that a designer could ship without rewriting.

Prompt 1: Infographic Brief Studio:

Reports → Create your own

Build a complete infographic for {topic}. Output:
1) The single key insight the infographic must communicate (one 
   sentence)
2) Information hierarchy (what the eye should see first, second, third)
3) Visual metaphor that maps the abstract idea to something concrete, 
   with the source-backed reason it fits
4) Layout type (vertical scroll, comparison, flowchart, stat-driven, 
   timeline) with the reasoning
5) Color logic (number of colors, what each color signals, accent rule)
6) Typography pairing (display font role, body font role)
7) The three things this infographic must NOT include

Hand this brief to Canva or a designer as-is. No further interpretation 
needed.

Customizable: {topic} → the AI concept you want to visualize.

Topic: Chain of Thought

Let’s see.

Prompt 2: Concept Infographic Studio:

Infographic → Customize

Visualize {topic} drawing on the strongest examples in the sources. 
Lead with one statistic from the sources as the focal point. Surround 
it with 4 supporting facts, each pulled directly from a different 
source. Skip generic AI explanations.

Customizable: {topic}.

My topic: Chain of Thought

Let’s see.

Prompt 3: Visual Hierarchy Audit Studio

Slide Deck → Create your own

Generate a slide deck teaching {topic} to a creator audience. For each 
slide: 1) Title (max 6 words), 2) Three bullets (max 10 words each), 
3) Speaker notes citing source location. Open with a hook slide. Close 
with a one-line takeaway. No filler slides.

Customizable: {topic} →

My topic: Chain of Thought

Prompt 4: Brand Voice Extract Studio

Video Overview → Create your own

Take the viewer on an immersive tour of how AI educators visually 
communicate complex ideas. Anchor each scene to a specific design move 
documented in the sources. Build to one signature visual that defines 
the field.

Here is the output.

Prompt 5: Comparison Infographic Studio

Infographic → Customize

Compare two approaches from the sources side by side. Pick the two 
techniques the sources contrast most often. Show 5 dimensions of 
comparison. For each dimension, give a one-phrase verdict and the 
source that backs it. End with one line on which approach the 
evidence favors.

Here is the output.

Prompt 6: Design Principle Infographic Studio

Infographic → Customize

Visualize the core design principles for AI educational content drawn 
from the sources. Lead with the single principle the sources cite most 
often as the focal point. Surround it with 6 supporting principles, 
each pulled from a different source. For each principle, include one 
8-word rule statement and the source it comes from.

Here is the output.

(I selected Chain-of-Thought for each topic on purpose. If you understand it even a little, just seeing it shows the power of NotebookLM.)

Next Steps

Pick one persona. Build the notebook tonight.

Do not try all four at once. The point of this post is the pattern, not the volume. Start with the persona closest to your daily work.

Steal the setup prompt as-is. Run two or three working prompts. See what NotebookLM gives back.

Once you have one persona working, the other three take 20 minutes each. The skeleton is the same. Sources, setup prompt, six prompts wired to Studio buttons.

Then start building your own. A copywriter persona for client work. A lawyer persona for contract review. A founder persona for investor questions. A coach persona for client homework. Same three-part structure, new voice.

The 24 prompts in this post are not the ceiling. They are the floor.

If you want more vault assets like this, full prompt libraries, agent setups, build walkthroughs, join the paid tier.

Steal these. Ship something this weekend.

Tell me what you built.

Everyone posts. Few convert. The ones who do convert follow a pattern most people never see, because they’re too busy writing their own Notes to study anyone else’s.

I spent a week reading Notes from 11 writers I respect →

Not the kind of reading where you scroll and double-tap.

The kind where you open the HTML, extract every Note, and ask what makes this person’s Notes recognizably theirs.

Eleven writers. All in the AI and tech space.

All actively posting. Some have 50K subscribers, some have 200.

Here’s what I found. Each writer has a Note DNA.

Similar to this →

A small set of formulas they run on repeat.

Once you see it, you can’t unsee it.

And once I saw it, I built something.

For every writer on this list, there’s a Claude Artifact at the end of their section.

Similar to this →

You type a topic, pick one of their structural templates, and the artifact produces a Note in that writer’s voice.

A Note that respects their rhythm, their openings, their quirks.

Similar to this →

Try one.

Then go read the real writer’s Notes and see the difference.

But first, let me explain the methodology.

The Methodology

We’ll do it in two simple steps.

Step 1- Data Collection

Go to the writer’s Substack page.

Let’s start with Wyndo, visit his page.

And scroll down 30 days. (Takes 5-6 minutes)

Now, right-click and click on “Save as”.

And save as wyndo-substack.html.

Step-2: Analyze with AI and Create an Artifact

Now let’s analyze with AI and build an artifact.

Open a new chat in Claude, attach this HTML, wyndo-substack, and paste this prompt.

You are analyzing a Substack writer's Notes to extract their writing DNA.
I have attached the HTML of their Notes page. Your job is to reverse-engineer
how they write so another person could produce Notes in the same voice.

Step 1. Extract the Notes from the HTML.
Step 2. Analyze opening patterns, sentence rhythm, structural formulas,
        topic territory, voice markers, and what they avoid.
Step 3. Produce a DNA card in a fixed format.
Step 4. Build a single-file HTML artifact that generates Notes in this
        writer's voice, using the Anthropic API.

Output all four in full. Do not skip steps.

And the analysis has started.

Let’s see how Wyndo writes such good Substack notes.

1. Wyndo — AI Maker

Wyndo writes about AI the way a carpenter talks about wood. He's interested in how it behaves, not how it looks on camera.

What I noticed: Wyndo treats prompts like products. He publishes them verbatim, in quotes, as if they’re the artifact.

The Extraction

The script pulled 203 Note containers out of the HTML.

After stripping Substack’s chrome, video controls, timestamps, and “Latest post” cards, 172 of them were real voice Notes.

The other 31 were podcast embeds and post promotions.

172 is a big enough sample to trust. Here’s what came out of it.

Wyndo’s Shape

Median length: 53 words.

Most common structure: 5 lines. 40% of his Notes are exactly five lines long. Another 14% are six.

Line length: 8 words on average. Almost half his lines are under 7 words. He treats every new line like a complete sentence. Capital letter to start, period to end.

Openers: 25% start with “I”. 20% start with “The”. Almost half of everything he writes begins with one of those two words.

Closers: A declarative principle. Something under 10 words that reads like a screenshot waiting to happen. He almost never ends on a question.

The 5-Line Skeleton

After reading 172 Notes, one pattern kept repeating. This is the shape:

L1  Hook with colon at the end
L2  A verbatim prompt in quotes
L3  Why it works — the mechanism
L4  The wider principle
L5  A maxim under 10 words

That’s it. 37% of his Notes use this exact structure.

Once you see it, you can’t unsee it.

What He Avoids

Em dashes. 172 Notes, 8 em dashes. He wants a clean copy that doesn’t look AI-generated.

Hashtags. At-mentions. Numbered lists. “Hot take” framing. Generic engagement bait like “DM me” or “comment below”.

He doesn’t screenshot his ChatGPT conversations. He quotes the prompts directly. The prompt is the object, not the AI response.

No self-deprecation. No false modesty.

He writes from earned authority, and the Notes read that way.

DNA Card

WRITER: Wyndo
PUBLICATION: AI Maker

CORE FORMULA:
Teach people to think with AI, not outsource to it.
Every Note is a portable upgrade: a prompt, a practice,
or a frame the reader can use tomorrow.

TOP 3 OPENING PATTERNS:
1. Colon setup — "The prompt that changed how I decide:"
2. First-person pivot — "I stopped doing X and started Y"
3. Twist reveal — "It's not about the tool. It's about..."

STRUCTURAL TEMPLATES:
1. Colon-Prompt: hook with colon, verbatim prompt,
   mechanism, principle, maxim
2. Pivot: "I stopped / started" or "Instead of / I"
3. Twist: thesis, "It's not X. It's Y.", portable restatement

VOICE MARKERS:
- Straight quotes (not curly) for prompts
- Every line capitalized, every line a full sentence
- Specifics over adjectives (four weeks, ten times, saves
  me hours — not "a lot" or "recently")

WHAT MAKES THIS WRITER DIFFERENT:
Wyndo is one of the only writers who publishes prompts
as the product. Not screenshots. Not paraphrases. The
actual prompt, in quotes, as the Note itself. That move
is his signature.

Try the Artifact

Pick a template. Type a topic. See what comes out.

Then go read Wyndo’s actual Notes and like&restack the ones you like, and compare.

Here is the link to the artifact, trained on Wyndo’s Notes.

Here's how it works.

2. Ilia Karelin — Prosper

Same steps I followed.

• Step 1: Go to Illia

• Step 2: Scroll down 1 month.

• Step 3: Save it as HTML.

• Step 4: Go to Claude.ai and paste the master prompt, which is:

You are analyzing a Substack writer's Notes to extract their writing DNA.
I have attached the HTML of their Notes page. Your job is to reverse-engineer
how they write so another person could produce Notes in the same voice.

Step 1. Extract the Notes from the HTML.
Step 2. Analyze opening patterns, sentence rhythm, structural formulas,
        topic territory, voice markers, and what they avoid.
Step 3. Produce a DNA card in a fixed format.
Step 4. Build a single-file HTML artifact that generates Notes in this
        writer's voice, using the Anthropic API.

Output all four in full. Do not skip steps.

Here is what it looks like.

And the analysis finished, let me summarize for you.

Ilia writes about AI tools.

His Notes are almost entirely about Claude, Cowork, Dispatch, and prompting.

He stays in his lane and sharpens it every week.

What I noticed: Ilia ends Notes with couplets that compound. “Output is temporary. Clarity compounds.” “Depth beats novelty.” 1 in 3 of his notes close with something you could screenshot and pin.

The Extraction

The script found 92 Note anchors in the HTML. After stripping Substack’s chrome and filtering out “new post” announcements, 71 original Notes remained. 11 were link broadcasts, not voice writing.

71 is a smaller corpus than Wyndo’s, but Ilia’s patterns are so repetitive that 71 is enough. Possibly more than enough.

Ilia’s Shape

Average length: 41 words. Shorter than Wyndo.

Lines: 4-5 per Note.

Here’s the detail that separates him from everyone else on this list. 32% of his sentences are 4 words or fewer.

That’s not a rhythm. That’s a signature.

The shape looks like this:

Short. Short. One longer line for texture. Short. Close shorter.

Fragments are on purpose. “Taste isn’t.” “Every time.” “Just one.” He treats a two-word sentence the same as a ten-word one.

The Negation Pivot

The move he runs more than any other is the negation pivot. “X isn’t Y. It’s Z.”

Examples from his actual Notes:

• “The best AI skill isn’t prompting. It’s editing.”

• “AI won’t fix a vague goal. It’ll help you pursue the wrong thing faster.”

• “Claude Code doesn’t make you a better developer. It makes your current thinking faster.”

25+ of his 71 Notes use this structure.

That’s 35%.

When you read him long enough, you can feel the pivot coming before it lands.

The Aphoristic Close

Ilia doesn’t end on a question. He ends on a couplet.

“Output is temporary. Clarity compounds.”

“Constraints beat aspirations every time.”

“Habit beats hype.”

“Depth beats novelty.”

“Better questions lead to better answers.”

The closing line is always shorter than the average sentence in the Note. It earns the last word by being the shortest.

What He Avoids

No em dashes. 10 across 71 Notes, so 1 every 7 Notes. Almost nothing.

No emojis. No hashtags. No exclamation spam.

No setup before the thesis. He starts with the point and lets the mechanism follow. Most writers do the opposite.

No hedges. “I think maybe” doesn’t exist in his writing. Neither does “kind of” or “sort of.”

No listicles. No name-dropping. No personal life except one “my dad” Note in the whole corpus.

DNA Card

WRITER: Ilia Karelin
PUBLICATION: Prosper

CORE FORMULA:
Calm expert posture. The tool is fine. Your thinking is
the variable. Every Note is a short, declarative
correction to how most people use AI.

TOP 3 OPENING PATTERNS:
1. Colon-hook — "The prompt that saves me from rabbit holes:"
2. Flat declarative — "Context is the skill."
3. Negation — "Bad AI outputs aren't frustrating anymore."

STRUCTURAL TEMPLATES:
1. Negation Pivot: "X isn't Y. It's Z." + mechanism + close
2. Stopped/Started: "I stopped X. Started Y. [aphorism]"
3. Most People / The Useful Move: default → reframe → consequence

VOICE MARKERS:
- Fragments on purpose (2-4 word sentences)
- Compounding couplets as closers
- Contractions always (doesn't, isn't, it's)

WHAT MAKES THIS WRITER DIFFERENT:
Ilia's closing line is a product. Most writers drift
to a conclusion. Ilia lands an aphorism that works on
its own, screenshot-ready, often more quotable than
the Note that carried it.

Try the Artifact

Pick a template. Type a topic. See what comes out.

Then go read Ilia’s actual Notes and like&restack the ones you like, and compare.

Here is the link to the artifact, trained on Ilia’s Notes.

Here's how it works.

You know how it works by now.

From now on, we’ll repeat everything.

3. Frank Andrade — Artificial Corner

Frank writes about AI.

He picked one tool and went deep.

84% of his Notes mention Claude by name.

What I noticed: Frank uses ✦ ☑ → as bullet glyphs instead of emojis. 21 instances across 45 Notes. Most writers reach for emoji when they want visual variety. Frank reaches for typography. The choice makes his Notes look like a printed handout, not a social post.

The Extraction

87 Note links in the MHTML. After deduplication and filtering to @thepycoach, 45 were Frank’s. The rest were restacks of other writers.

45 is the smallest corpus in this analysis. But Frank is so consistent that 45 is enough. He’s running the same playbook on repeat, and the playbook reveals itself fast.

Frank’s Shape

Average: 2.9 sentences per Note before he transitions into a list.

First-line word count: 6-12 words. Punchy hooks, never longer.

Sentences over 15 words almost don’t exist. He keeps it tight, then opens up into structured lists.

The rhythm is short hook, two-beat reversal, then bulleted payload.

The Two-Beat Reversal

Frank’s signature opening rhythm is a snare-then-kick.

“Most people burn through Claude’s limits by lunch. Not because they use it too much. Because they use it wrong.”

“You don’t need a better prompt. You need the right Claude.”

[wrong assumption]. [actual answer].

Short declarative. Correction. The correction is always shorter than the setup, and it always carries the surprise. He runs this rhythm at the top of his Notes the way Wyndo runs his colon-prompt structure — as a recognizable opener.

The Step Stack

His most repeated structural template:

1. Hook

2. “Just follow these steps:”

3. Step 1: [action]

4. Step 2: [action]

5. CTA or link

About 10 of his 45 Notes use this exact structure. The Note reads like documentation. No prose, no stories, no personal asides. Just the steps.

His second template is the prompt library:

“Start with these N prompts:

#1 [Name]

‘[the prompt itself in quotes]’”

The prompts are quoted verbatim, like Wyndo. But Frank wraps them in a numbered list with explicit naming. Wyndo publishes one prompt per Note. Frank publishes 8 in a single post.

“Better Than 99% of People”

A formula Frank uses 3 times in 45 Notes that I haven’t seen anywhere else on this list:

“How to set up Claude better than 99% of people.”

“Set up Claude Cowork better than 99% of people.”

The hook positions the reader as someone about to leapfrog the crowd. It’s a confidence promise. The body of the Note then has to deliver — and Frank delivers with bullet lists of specific settings, prompts, or features.

This is the kind of move most writers avoid because it sounds boastful. Frank uses it because it sounds useful. The difference is whether the Note that follows actually backs the claim. His do.

What He Avoids

Em dashes. 6 across 45 Notes. Mostly in compound phrases, not for drama. Same household rule as everyone else on this list.

No personal life. No family. No location. No politics. No hot takes about other creators.

No corporate language. No “excited to share,” no “humbled,” no “thrilled.”

No hedging. No “I think” or “in my opinion.” He asserts.

No cliffhanger questions. He never closes with “what do you think?” or “agree?”. He closes with a CTA, a link, or a “see more” cliff that the platform inserts after the truncation point.

What He Does That Others Don’t

Frank front-loads. He puts the value above the fold and parks the payoff below “see more.” The reader has to click to expand or click through to artificialcorner.com to get the rest. The Note is the hook. The article is the deliverable.

Most writers on this list treat Notes as the artifact itself. Frank treats them as the door. That distinction shows up in his closings — they’re almost never standalone aphorisms or punchlines. They’re transitions to somewhere else.

He also uses parenthetical asides for stage direction:

“(read them in this order)”

“(free guides included)”

“(link to download all the skills at the end 👇)”

These are small but they work. The asides feel like a teacher leaning in to tell you something extra. They make a list-heavy Note feel personal without him having to write a personal sentence.

DNA Card

WRITER: Frank Andrade
PUBLICATION: Artificial Corner

CORE FORMULA:
Practical Claude educator for non-technical professionals.
Hook with a two-beat reversal, then a structured list,
then a CTA to artificialcorner.com. Notes are the door,
articles are the room.

TOP 3 OPENING PATTERNS:
1. Status drop — "Claude Opus 4.7 just dropped."
2. Diagnosis — "You're using AI wrong."
3. Authority claim — "The best Claude prompt I use is..."

STRUCTURAL TEMPLATES:
1. Step Stack: hook, "Just follow these steps:",
   numbered actions, CTA
2. Prompt Library: hook, "Start with these N prompts:",
   #1 [Name], "[verbatim prompt]"
3. Bullet Showcase: hook, ✦ feature, ✦ feature, link

VOICE MARKERS:
- ✦ ☑ → as bullet glyphs instead of emoji
- "X better than 99% of people" formula
- Year-stamping ("in 2026") for currency
- Parenthetical asides as stage direction
- Numbers everywhere (80%, 99%, $20/month)

WHAT MAKES THIS WRITER DIFFERENT:
Frank treats Notes as the door, not the room. Every
other writer on this list closes with something
self-contained — a maxim, a punchline, a verdict.
Frank closes with a link. The Note's job is to
qualify the click. The article does the teaching.

Try the Artifact

Pick a template. Type a topic. See what comes out.

Then go read Frank’s actual Notes and like&restack the ones you like and compare.

Here is the link to the artifact, trained on Frank’s Notes.

Here’s how it works.

4. Joel Salinas — Leadership in Change

Joel writes for leaders.

His Notes sound like a consultant who stopped billing by the hour and started thinking out loud.

What I noticed: Joel uses the word “leaders” the way most tech writers use “founders.”

It’s his default reader.

Every time he writes “Most leaders...”, he’s signaling who the Note is not for, and that clarity is what makes it land.

The Extraction

79 original Notes after filtering out restacks. Restacks were excluded because they carry someone else’s voice.

What’s left is Joel’s own writing, across strategy, governance, prompting craft, and team enablement.

Joel’s Shape

Median length: 28 words. Shorter than both Wyndo and Ilia.

But the distribution is bimodal. Half his Notes are under 20 words. The other half are 50-90 words. He skips the middle. You’re either getting a micro-insight or a thought-leadership beat. Rarely the in-between.

Average sentence: 8 words. 23% of his sentences are 4 words or fewer.

Triads show up constantly:

“Prompt. Spin. Regenerate. Hope.”

“Emails. Slack. Calendar.”

Three short declaratives stacked, then one longer line that prescribes. That’s the engine.

Diagnose → Prescribe

Joel’s signature structure reads like a doctor’s note. State the problem. Give the evidence. Land the fix.

Example from his corpus:

“Most strategic failures don’t happen because leaders lack intelligence.

They happen because leaders can’t test assumptions fast enough.

The decision window is weeks. The consultant timeline is months.

Twenty minutes of the right AI questions can compress months of expert consultation into something actionable.”

16 of his 79 Notes open with this collective-failure diagnosis. “Most leaders...”, “Most teams...”, “Most strategic...”. He names the failure first, then prescribes.

The “Your X” Stack

When Joel isn’t diagnosing, he’s stacking “Your” declaratives:

“Your team is using AI. Your competitors know it. Your data policies should reflect it.”

Three short sentences, each a fact the reader can’t argue with, ending in a call to act. It reads like a memo that skipped the pleasantries.

What He Avoids

No em dashes. One instance in 79 Notes. That’s a habit, not an accident.

No hype vocabulary. No “game-changer,” no “10x,” no “unlock,” no “supercharge.”

No bullets. No numbered lists. He stacks sentences with line breaks instead.

No hedging. No “I think maybe this might possibly...”

No invented stats. When he gives a number, it comes from his own workflow, not a made-up benchmark.

What He Does That Others Don’t

Joel gives credit generously. He tags collaborators. He praises other writers. He sometimes closes on a soft beat — “:)” or “Love seeing that!” — after a sharp business Note.

This is rare in the AI-tech space. Most writers in the niche perform seriousness. Joel doesn’t perform. He’s serious when the point demands it and warm when the moment allows it.

That mix is part of his DNA.

DNA Card

WRITER: Joel Salinas
PUBLICATION: Leadership in Change

CORE FORMULA:
Diagnose what most leaders do wrong. Give 2-4 beats
of evidence. Land a short prescription the reader
can act on Monday morning.

TOP 3 OPENING PATTERNS:
1. Diagnostic — "Most leaders..." / "Most teams..."
2. "Your" stack — "Your team is using AI."
3. Reader question — "Do you trust AI more for
   research, drafting, or reviewing, and why?"

STRUCTURAL TEMPLATES:
1. Diagnose → Prescribe: failure, evidence, corrective
2. "Your" Stack: three "Your X" sentences, then payoff
3. Reframe → Payoff: common belief, flip, quotable line

VOICE MARKERS:
- "Leaders" as default reader noun
- Triads of short declaratives
- Ellipsis (...) for pause or teaser in ~30% of Notes
- Occasional warm closer after a sharp point

WHAT MAKES THIS WRITER DIFFERENT:
Joel is writing for an audience that most AI writers
ignore — executives and managers, not builders. His
tone reflects it. He sounds like someone who's been
in the boardroom, not someone who's watched a video
about the boardroom.

Try the Artifact

Pick a template. Type a topic. See what comes out.

Then go read Joel’s actual Notes and like&restack the ones you like and compare.

Here is the link to the artifact, trained on Joel’s Notes.

Here’s how it works.

5. Karo — Product with Attitude

Karo is an AI Product Manager.

Her Notes sound like a senior PM thinking out loud, calm, committed, and uninterested in hype.

What I noticed: Karo writes aphorisms that stand alone. One line. One period. No setup, no payoff, no explanation.

The Extraction

95 Notes by Karo in the HTML.

After separating restacks, link shares, and quote echoes from original thoughts, 56 original-thought Notes remained.

The rest were her amplifying other writers or sharing her own links.

56 is on the smaller side, but her patterns are tight enough that the signal is clean.

Karo’s Shape

Median sentence: 8 words. Same as Joel.

But 36% of her sentences are 6 words or fewer. That’s her rhythm, short stacks, then one shorter line that lands.

She rarely crosses 20 words in a sentence.

When she does, it’s deliberate.

The long sentence sets up a short punchline.

Here’s a real example of the rhythm:

“My ideas don’t arrive in order. They show up all at once. AI has been an incredible thinking partner for me, but only after I learned how to use it for better questions, not faster answers.”

Three short. One long. The long one is where the payoff hides.

The Colon Reveal

23 colons across 56 Notes.

That’s almost one every two Notes. She uses the colon to frame before delivering.

“Before AI: Ideas were easy, building was hard. After AI: Building is easy, choosing the right idea is hard.”

The colon is the switch. The left side sets the frame. Right side flips it.

What She Avoids

No em dashes. 1 instance across 56 Notes. She uses periods where most writers reach for em dashes.

No semicolons. Zero.

No superlatives. No “incredible,” no “insane,” no “mind-blowing.” She could be writing about a genuine breakthrough and still refuse the marketing vocabulary.

No “Here’s what nobody tells you...” framing. No rhetorical self-questioning. No hedges like “I think maybe” or “arguably.”

No redemption arcs. She doesn’t write “I used to do X, then I realized Y.” When she states a position, she just states it.

DNA Card

WRITER: Karo
PUBLICATION: Product with Attitude

CORE FORMULA:
Short, committed declaratives about critical AI
literacy. Aphorisms that stand alone. Colons that
flip. Occasional personal build stories.

TOP 3 OPENING PATTERNS:
1. Cold metaphor — "X is the fast food of Y."
2. Binary frame — "Before AI: ... After AI: ..."
3. Flat negation — "There's no such thing as X."

STRUCTURAL TEMPLATES:
1. Colon Reveal: frame, then payoff — "X: Y."
2. Aphorism-Only: one line, one period, no elaboration
3. Setup-Pivot-Land: 3 short sentences, last hits hardest

VOICE MARKERS:
- Curly quotes, never straight
- 1 emoji maximum per Note (🤗 for warmth, 😂 for
  self-deprecation)
- Commits to views — no hedging
- Names collaborators by full handle when tagging

WHAT MAKES THIS WRITER DIFFERENT:
Karo is the only writer on this list who posts
aphorisms as complete Notes. One line. One period.
No context. The aphorism is the product. Everyone
else wraps their best line in an explanation. She
just posts the line.

Try the Artifact

Pick a template. Type a topic. See what comes out.

Then go read Karo’s actual Notes and like&restack the ones you like and compare.

Here is the link to the artifact, trained on Karo’s Notes.

Here’s how it works.

6. Karen Spinner — Wondering About AI

Karen builds a product called CarouselBot.

She writes about what happens while she’s building it.

What I noticed: Karen is the only writer on this list who runs experiments before offering opinions. Most AI writers take a position and find evidence.

Karen takes 500 API calls and lets the result decide what she thinks.

The Extraction

187 feed units in the MHTML.

After separating her own Notes from restacks and article captions, 107 were hers. 35 of those were long-form prose Notes where her real voice lives.

The rest were captions on restacked articles or short endorsements.

35 long Notes is a smaller sample than the others, but Karen’s voice is so consistent that it doesn’t take a big corpus to read her.

Karen’s Shape

Average sentence: 15.6 words. Longer than everyone else so far.

But 20% of her sentences are 5 words or fewer. Those short ones are where the jokes live.

Her rhythm is long → long → short punch. Example from her corpus:

“At one point, Claude told me the code was compiled and ready to test. It wasn’t. 😂”

Two sentences. One sets up. “It wasn’t.” closes it. The 3-word sentence is the whole joke.

The Experiment → Finding → Recommendation

Karen’s signature structure is a lab report compressed into a Note.

1. Here’s what I ran

2. Here’s what I found

3. Here’s what I’m doing differently now

Real example from her corpus:

“Inspired by X research, I ran 250 API calls... Interestingly enough, telling Claude about its constraints produced 100% correctness. Here’s what I’m doing differently now:”

That “Here’s what I’m doing differently now:” is her verbal tic. It shows up across multiple Notes. Other writers would write “So the lesson is...” Karen writes it as an action she already took.

The Bug Story

Her third move is the confession Note. Something broke; here’s what she learned.

“I was exploring Postgres Neon... accidentally clicked the wrong button... which nuked the production database! 🙀 This was a good reminder that sometimes you need MORE friction, not less.”

Most AI writers demonstrate expertise. Karen performs the opposite. She tells you what she broke before she tells you what she figured out. The authority comes from surviving the mistake, not avoiding it.

What She Avoids

No em dashes. 2 in 107 Notes. Same habit as everyone else on this list so far.

No LinkedIn openers. No “Hot take:”, no “Unpopular opinion:”, no “Most people get X wrong.”

No empty superlatives. “Amazing” and “incredible” don’t appear in her writing.

No generic takes. If she has an opinion about a model, it’s attached to a specific version and a specific test. “Opus 4.6 does X.” Not “Claude does X.”

No hustle vocabulary. No “crushing it,” no “shipping hard.” She ships, but she doesn’t brand it.

What She Does That Others Don’t

Karen names names. When she boosts a peer, she uses their full name and what they specifically did. Not “great post from @someone” — “Thanks to Daria Cupareanu for walking through the exact prompt she used to...”

She uses emojis sparingly and tends to use them at the end of a clause. 🤔 for musing. 😂 for self-roasting. 🙀 when something broke badly. The emoji budget is strict.

She trails off mid-sentence with an ellipsis. “So…I’m writing a technical report.” “It’s not a bad design. It’s just…Claude design.” The ellipsis is her hesitation on purpose.

DNA Card

WRITER: Karen Spinner
PUBLICATION: Wondering About AI

CORE FORMULA:
Builder reporting from the lab, not commentator
offering opinions. Every Note is attached to a
specific experiment, a specific bug, or a specific
model version. No abstractions without evidence.

TOP 3 OPENING PATTERNS:
1. First-person experiment — "I ran 250 API calls to..."
2. Declarative + pivot — "Claude Design is powerful. But..."
3. Bug story — "I was exploring X and accidentally..."

STRUCTURAL TEMPLATES:
1. Experiment → Finding → "Here's what I'm doing differently now:"
2. Received wisdom → "But" → counter-evidence → so-what
3. Bug story → what I did → lesson

VOICE MARKERS:
- One emoji per Note, always at end of clause
- Mid-sentence trailing ellipsis (So…I'm thinking)
- Pivots with "But" or "Then" — never "However"
- Full-name credit when boosting peers

WHAT MAKES THIS WRITER DIFFERENT:
Karen is the only writer on this list who runs an
experiment before forming an opinion. Most writers
start with a position and find support. Karen takes
500 API calls and lets the data tell her what the
Note is about. The experiment is the argument.

Try the Artifact

Pick a template. Type a topic. See what comes out.

Then go read Karen’s actual Notes and like&restack the ones you like and compare.

Here is the link to the artifact, trained on Karen’s Notes.

Here’s how it works.

7. Jenny — Build to Launch

Jenny writes about Claude Code, MCP servers, prompts, and what actually works when you’re building something real.

What I noticed: Jenny is the only writer on this list who uses em dashes on purpose. Every other writer avoids them. Jenny uses 89 across 105 Notes. That’s not an accident. It’s a stylistic choice she’s committed to.

The Extraction

131 Note bodies in the MHTML. After filtering by author permalink, 112 were Jenny’s. After dedup, 105 unique Notes.

The decoding had a quirk, quoted-printable encoding mangled the emojis on the first pass. Turns out there weren’t any.

Jenny doesn’t use emojis in her teaching Notes. That “mangled emoji” was just a tree character (↳) she uses as a list marker.

Jenny’s Shape

Average length: 54 words. Hard-capped under 100. The ceiling is deliberate.

Modal structure: 4 paragraphs.

Median sentence: 7 words. 37% of her sentences are 5 words or fewer.

Her rhythm is long → short → short. A statement, then two fragments that land. You can feel her picking a complete sentence, then deciding the last two beats don’t need one.

The “I + Past Tense + Number” Opener

34% of Jenny’s Notes open with her doing something specific, with a number attached.

• “I installed 11 plugins...”

• “I tested this with 40 prompts...”

• “I killed my bookmarks folder...”

• “I kept 3 out of 18 cron jobs...”

The number does the work. It signals this isn’t advice. It’s a report.

Another 19% open with a quantifier, “Every,” “Most,” “Everyone,” or a count. Her openings are built on specificity.

She almost never opens with a question. The one thing she doesn’t do is ask you something to get you engaged.

The “Most People X. But These N Moves” Structure

Her signature teaching structure looks like this:

1. Most people do X wrong

2. These N specific moves fixed it for me

3. Each move is one em-dash bolded label plus one sentence

4. Optional follow-up line with ↳ as the marker

It reads like documentation written by someone who’s tired of answering the same question in DMs.

The Em Dash Habit

Here’s what separates Jenny from everyone else in this analysis.

Wyndo: 8 em dashes across 172 Notes. Ilia: 10 across 71. Joel: 1 across 79. Karo: 1 across 56. Karen: 2 across 107.

Jenny: 89 across 105.

That’s almost one per Note. Em dashes are her primary device for mid-sentence pivots. Where Ilia uses a period, Jenny uses an em dash. Where Karo uses a colon, Jenny uses an em dash.

If you’re copying Jenny’s voice, the em dashes are not optional.

Closings That Reframe

Jenny doesn’t summarize at the end. She reframes.

“It’s a clarity problem.”

“The old prompt got agreement. The new one gets a bug report.”

“The test at the end isn’t extra. It’s the whole point.”

The closing line recontextualizes everything above it. A Note that sounded like a list of tactics suddenly becomes a Note about one principle. The last sentence is where she earns the read.

What She Avoids

No emojis in teaching Notes. Zero.

No exclamation marks. Her tone stays flat when she’s teaching.

No hype vocabulary. No “game-changer,” “leverage,” “unlock,” “secret,” or “hack.” Not once in 105 Notes.

No ALL CAPS. No “guys” or “folks” as audience address.

No hashtags. No DM asks.

DNA Card

WRITER: Jenny Ouyang
PUBLICATION: Build to Launch

CORE FORMULA:
Senior-engineer log entries. Every teaching Note is
anchored in a specific number she ran herself.
Hard-capped under 100 words, typically 4 paragraphs,
with em dashes doing the pivots.

TOP 3 OPENING PATTERNS:
1. "I + past tense + specific number"
   — "I installed 11 plugins..."
2. Quantifier lead — "Every," "Most," "Everyone"
3. Thing-that-worked — "The thing that made AI
   actually useful for Y..."

STRUCTURAL TEMPLATES:
1. "Most people X. But these N moves..." with em-dash
   bolded labels and ↳ follow-ups
2. "I had X, Y, Z. Then I stopped." confession structure
3. Aphorism-to-evidence: short claim, then the numbers
   that justify it

VOICE MARKERS:
- Em dashes as primary mid-sentence pivot (89 in 105)
- ↳ tree character for list follow-ups
- Specific numbers in almost every teaching Note
- Closings that reframe, not summarize

WHAT MAKES THIS WRITER DIFFERENT:
Jenny runs the experiment and keeps the receipt.
Her Notes read like a log book, not a blog. The
numbers she cites — 11 plugins kept 3, 18 cron
jobs, 40 bookmarks — make the teaching credible
in a way opinion-based Notes can't touch.

Try the Artifact

Pick a template. Type a topic. See what comes out.

Then go read Jenny’s actual Notes and like&restack the ones you like and compare.

Here is the link to the artifact, trained on Jenny’s Notes.

Here is what it looks like.

8. Daria — AI Blew My Mind

Daria writes about Claude the way someone writes about a tool they actually use every day.

Her publication name is a promise. AI Blew My Mind. And 16 times across her corpus, she says exactly that phrase.

What I noticed: Daria drops one-line punchlines after 2-3 setup paragraphs, and the punchline is almost always under 10 words. “Just move.” “This is why I do what I do.” The setup builds weight. The punchline releases it in one breath.

The Extraction

189 feed items in the MHTML. After filtering 29 restacks and 4 post-share cards, 156 original Notes remained. Total corpus: 7,735 words.

The top 30 by likes were saved separately. Those are the Notes where her voice lands hardest.

Daria’s Shape

Median sentence: 7 words. Same as Joel. Same as Ilia.

But here’s what separates her: 18% of her sentences are 3 words or fewer. That’s extreme. Even Ilia at 32% fragments doesn’t push this short.

Fragments are her rhythm reset. Long sentence. Shorter sentence. Fragment. Long sentence again.

46% of her paragraphs are single-sentence. Almost half of everything she writes is one sentence standing alone on its own line.

The Setup → Punchline

Daria’s signature structure is 2-3 paragraphs of setup, then a one-line closer under 10 words.

33% of her Notes end this way. It’s the most consistent structural signal in her corpus.

The setup does the argument. The punchline does the release.

[setup]

Just move.

[setup]

This is why I do what I do.

The punchline is always shorter than any sentence in the setup. That contrast is what makes it land.

“Not X. But Y.”

Her second signature move is the negation flip, but with a twist. Where Ilia uses “X isn’t Y. It’s Z.” Daria uses “X isn’t Y. But Y.” — making the “but” do the pivot work.

“AI isn’t built to know things. It’s built to relate things to each other.”

Same mechanical shape as Ilia. Different cadence. Ilia’s version feels like a verdict. Daria’s feels like a realization.

Anaphora

15% of her Notes use anaphora — the same 2-word opener repeated across consecutive sentences.

“That’s why it gets the big picture right.

That’s why it can give you a solid business plan outline...”

“That’s why” is her favorite causal bridge. She uses it to stack evidence without ever having to say “furthermore” or “additionally.” The rhythm does the connecting.

What She Avoids

No em dashes. 5 across 7,735 words. Consistent with the pattern across this list.

No semicolons. Zero.

No hype vocabulary. “Amazing,” “incredible,” “revolutionary,” “game-changing” don’t appear once.

No thread formatting. No “1/”, no “🧵”, no “a thread:”.

No meta-openers. No “Let me tell you...” No “In this post...”

No hedging. “Maybe,” “perhaps,” “I think” are rare.

What She Does That Others Don’t

Daria drops real quotes. 24% of her Notes contain dialogue — from readers, clients, her mom, or herself. The quote is the evidence.

“AI just blew my mom’s mind.” That’s a full Note opener.

Other writers paraphrase what people say to them. Daria just puts it in double quotes.

She also writes to “you” more than to “leaders” (Joel) or “I” (Wyndo). Second-person is 18% of her openings. She addresses the reader directly and often — “You don’t figure out AI and then start using it.”

DNA Card

WRITER: Daria Cupareanu
PUBLICATION: AI Blew My Mind

CORE FORMULA:
Practitioner reporting, not pundit speculation. Every
Note comes from the chair she sits in when she ships.
Setup in paragraphs, release in a one-line punchline.

TOP 3 OPENING PATTERNS:
1. Declarative fact — "AI isn't built to know things."
2. Second-person directive — "You don't figure out AI
   and then start using it."
3. Contrarian setup — "Everyone wants to keep up with AI."

STRUCTURAL TEMPLATES:
1. Setup → Punchline: 2-3 paragraphs, one-line close
   under 10 words
2. Not X. But Y.: negation flip with "but" pivot
3. Anaphora: "That's why..." repeated across 2-3
   consecutive sentences

VOICE MARKERS:
- 18% of sentences are 3 words or fewer
- 46% of paragraphs are single-sentence
- Double quotes for dialogue or sarcastic framing
- "That's why..." as the causal bridge
- One emoji max, used sparingly (🥲 😄 🤭)

WHAT MAKES THIS WRITER DIFFERENT:
Daria is the only writer on this list who quotes
the people around her. Her mom. Her clients. Her
readers. 24% of her Notes contain real dialogue in
double quotes. Everyone else paraphrases. She just
copies the line and puts it on the page.

Try the Artifact

Pick a template. Type a topic. See what comes out.

Then go read Daria’s actual Notes and like&restack the ones you like, and compare.

Here is the link to the artifact, trained on Daria’s Notes.

Here is what it looks like.

9. Sam Illingworth — Slow AI

Sam is a professor and poet in Edinburgh. He writes about AI the way a good political journalist writes about power.

What I noticed: Sam is the only writer on this list whose subject isn’t AI craft. It’s AI power. Who builds it, who pays for it, who stays invisible in the coverage. Every Note is a press briefing with the press release removed.

The Extraction

117 unique Notes after deduplication. Average 231 words, median 71. The distribution is bimodal — short announcements and shoutouts on one side, longer commentary on the other.

The voice lives in the commentary half. The announcements are functional. The shoutouts follow a template he reuses almost word-for-word. What’s distinct is everything in between.

Sam’s Openings

Four moves, in order of frequency.

Statistic-as-sentence (37%). A hard number, planted as a standalone declaration.

“Grok 4’s training emissions reached 72,816 tonnes of CO2 equivalent.”

“Three in four US teens have already used an AI companion.”

Named actor + action + date (30%). A specific company, official, or institution doing a specific thing on a specific date.

“Mike Krieger sat on Figma’s board while Anthropic built a Figma rival.”

“Coventry City Council just renewed its Palantir contract for £750,000.”

Reframing aphorism (20%). A sentence that looks like a claim but is actually a trap being set.

“The company that built tools to watch everyone else is now watching its own staff.”

“When respected artists adopt AI, we call it evolution. When students do, we call it cheating.”

Imperative or question (13%). “Stop being polite to AI.” “Who owns the language of AI?”

None of these hedge. No “I think,” no “perhaps.” Cold open, every time.

Fact / Turn / Verdict

Sam’s signature structure is a three-beat move.

1. Open with a cold fact

2. Introduce the overlooked second fact that reframes the first

3. Close with a one-line verdict that names the mechanism, not the emotion

The Anthropic pricing Note is the clearest example:

“Anthropic announced last week that Claude Opus 4.7 costs the same per token as its predecessor.

They did not mention that the new tokenizer uses up to 35% more tokens to process the same input.

The per-token price is unchanged. Your bill is not.”

Three beats. The verdict isn’t an opinion. It’s the frame being named. “Per-token price is unchanged. Your bill is not.” That’s how a journalist closes a story when the story is really about how stories are told.

Clipped Parallelism

Sam’s rhythmic fingerprint is parallel short sentences, usually in threes.

“Not a forecast. Not a white paper. Not a press release.”

“No regulator. No court. No parliament.”

“Nobody to call. Nobody to fire.”

“The public pays for the infrastructure. The compute leaves the country. The reactors stay for decades.”

When he wants the reader to feel a pattern, he makes them hear it. Three identical structures. Each one shorter than a tweet. The rhythm carries the argument.

Naming the Frame

Here’s the move that separates Sam from every other writer on this list. He doesn’t just report what the coverage says. He describes how coverage works.

“The spending is reported as investment. The consequences are reported as externalities.”

“This is what a pricing-unchanged headline looks like when one number stays fixed and the other quietly moves.”

Most AI writers argue with the conclusions in industry headlines. Sam argues with the sentence structure. He’s writing about how the language itself hides the thing that matters.

What He Avoids

No em dashes. 2 across 117 Notes.

No emojis, except occasional 🚨 for corrections.

No exclamation marks in commentary Notes.

No hedging. “I think,” “in my opinion,” “personally” don’t appear. The evidence carries the claim.

No hype vocabulary. “Game-changer,” “revolutionary,” “insane” appear once across the corpus, and that instance is quoting someone else.

No listicles, no “1/” threads, no hook-then-thread formatting.

No speculation without a source. Every assertion comes with a name, a date, or a citation baked into the sentence.

What He Does That Others Don’t

Sam names people. Not “a Big Tech CEO” — Alex Karp, Mike Krieger, Rachael Maskell. Not “a study” — “Stanford’s 2026 AI Index,” “Common Sense Media’s risk assessment.” Names make the Note unignorable.

He uses British spelling. Behavioural. Summarise. Labour. Centre. Dates as “16 April,” not “April 16th.” Currency with the symbol up front: £330m, not “330 million pounds.”

He uses single quotes to scare-quote the industry’s own words. ‘Research preview.’ ‘Scalable oversight.’ ‘AI redundancy washing.’ The quotes say this is what they’re calling it, not what it is.

DNA Card

WRITER: Sam Illingworth
PUBLICATION: Slow AI

CORE FORMULA:
Journalist covering AI power, not AI craft. Every
commentary Note is built on named actors, specific
dates, and hard numbers. The verdict names the
frame, not the feeling.

TOP 3 OPENING PATTERNS:
1. Cold statistic — "72,816 tonnes of CO2 equivalent."
2. Named actor + action + date — "Coventry Council
   renewed its Palantir contract for £750,000."
3. Reframing aphorism — "The company that built tools
   to watch everyone is now watching its own staff."

STRUCTURAL TEMPLATES:
1. Fact / Turn / Verdict: cold fact, overlooked
   second fact, one-line mechanism-naming close
2. Stat / Mechanism / Moral cost: headline number,
   how it works, cost the coverage hid
3. Category swap: "When X does it we call it A.
   When Y does it we call it B. Same tool, different
   permission structure."

VOICE MARKERS:
- Clipped parallelism in threes
- British spelling and date format
- Single quotes for the industry's own words
- Sources named in-sentence, not appended

WHAT MAKES THIS WRITER DIFFERENT:
Sam doesn't argue with what industry coverage says.
He argues with how it says it. The subject of most
Notes is the sentence structure of the press release,
not the product in the press release. No other writer
on this list works this register.

Try the Artifact

Pick a template. Type a topic. See what comes out.

Then go read Sam’s actual Notes and like&restack the ones you like and compare.

Here is the link to the artifact, trained on Sam’s Notes.

Here’s what it looks like.

10. Yevheniya — Mother Using AI

Full disclosure: Yevheniya is my wife and the co-founder of Learn With Me AI. I ran her Notes through the same process as everyone else on this list.

The DNA card came back with patterns I didn’t fully see, even after reading her Notes for a year.

What I noticed: Yevheniya opens Notes with “AI won’t,” “AI didn’t,” or “AI doesn’t” 16 times across 99 Notes. That’s 16%. She sets up a fear or expectation about AI, then the second sentence reverses it. The reversal is her entire formula.

The Extraction

174 Note links in the MHTML. After deduplication, 106 unique Notes. 99 were originals. 7 were restacks of my own Notes, which I filtered out.

Median length: 25 words. Shorter than any other writer on this list.

Yevheniya’s Shape

Median sentence: 6 words. 48% of her sentences are 5 words or fewer.

Her rhythm is short-short-short-longer-short. Sometimes short-shorter-shortest. Fragments are fair game.

Examples from her corpus:

“3 hours a week back.”

“Automatically.”

“On repeat.”

“Different question. Different life.”

These aren’t decorative. They’re landings. A two-word sentence on its own line is where the Note earns its keep.

The Reversal

Her signature structure is the AI-reversal. 16 Notes open with it.

Sentence 1: “AI won’t / didn’t / doesn’t [feared outcome].” Sentence 2: “It [unexpected outcome].”

Real example:

“AI didn’t give me more time. It took away my favorite excuse.”

The reversal lives entirely in sentence 2. “It” does the work. 8 of her 16 reversal openings pivot with “It” as the subject of the second sentence.

This is the tightest structural signal I saw in any of the 9 writers analyzed so far. Nobody else runs a two-sentence setup-flip this often.

The Husband Story

8+ Notes open with “My husband [past-tense verb]...” Her second signature.

“My husband said ‘I’ll build a machine that writes SEO articles while I sleep.’ I said good luck with that. Now it scrapes Google, finds keyword gaps, and writes articles.”

Three beats. Claim. Skeptical narrator. Punchline.

She is always the skeptic. The husband is always the protagonist-builder. This frame lets her write about AI builds from the outside, which is a rare voice in the AI-tools niche. Most writers who publish builds are the builders. She’s the one watching someone build and asking whether it actually works.

Naked Numbers

Yevheniya uses numbers as punchlines. No setup. No “approximately.” No “around.”

“56% more.” “3 hours a week back.” “150K+ stars in 4 days.” “2,000 developers.” “$2,000 project.”

The number arrives on its own line or at the end of a sentence, stripped of hedges. The specificity makes the claim unarguable.

What She Avoids

No em dashes. Zero across 99 Notes.

No exclamation marks in AI-tools Notes. The 4 exclamations in her corpus all come from the earlier pregnancy cluster, where her voice was warmer.

No emojis in the AI-tools Notes. They show up in pregnancy Notes. Different context, different rules.

No hedging. “Maybe,” “I think,” “kind of” don’t appear in her declarations.

No rhetorical questions at the open. Questions show up later in a Note, if at all.

No corporate jargon. No “leverage,” “synergy,” or “democratize.”

What She Does That Others Don’t

Yevheniya is the only writer on this list who uses a recurring character other than herself. “My husband” shows up in 8+ Notes as a protagonist she narrates around.

This solves a problem that most AI-tools writers run into. When every Note is “I did X,” the voice gets self-centered fast. When some Notes are “someone close to me did X and I watched it work,” the voice becomes observational. Same topic, different chair.

She also uses “you” differently from the others. “You” doesn’t appear at the top of her Notes. It shows up mid-way, only to land the lesson. The Note is about her, or about me, or about a news beat — and then “you” arrives at the end to take the takeaway home.

DNA Card

WRITER: Yevheniya
PUBLICATION: Mother Using AI

CORE FORMULA:
Short, declarative reversals. Sentence 1 sets up a
fear or expectation. Sentence 2 flips it. Numbers
arrive naked. Fragments land the verdict.

TOP 3 OPENING PATTERNS:
1. AI won't / didn't / doesn't + reversal in sentence 2
2. "My husband + past-tense verb" — the husband story
3. "Most people X. I Y." — differentiation opener

STRUCTURAL TEMPLATES:
1. The Reversal: feared outcome, then "It [unexpected]"
2. Husband Story: claim, skeptical narrator, punchline
3. Before / After: "I used to X. Now Y."

VOICE MARKERS:
- 48% of sentences are 5 words or fewer
- "It" as pivot subject in sentence 2
- Numbers arrive naked — no "approximately"
- Parallel couplets as closers ("The tool is free.
  The X isn't.")

WHAT MAKES THIS WRITER DIFFERENT:
Yevheniya is the only writer on this list who writes
from the narrator's chair, not the builder's. Her
husband is the protagonist-builder in 8+ Notes. She
is the skeptic watching, then reporting what actually
worked. No other writer in the AI-tools niche uses
this observational frame.

Try the Artifact

Pick a template. Type a topic. See what comes out.

Then go read Yevheniya’s actual Notes and like&restack the ones you like and compare.

Here is the link to the artifact, trained on Yevheniya’s Notes.

Here’s what it looks like.

11. Gencay — LearnAIWithMe

The narrator goes last.

I ran the same prompt on my own Notes. If the Claude Artifact at the end of this section produced notes my wife couldn’t tell apart from mine, the method works.

What I noticed: I open Notes with obituaries more often than I realized. “R.I.P [X]. R.I.P [Y]. R.I.P [Z].” 12 times across 148 Notes. I didn’t know I was doing this on repeat. The DNA card showed me a pattern I’d been using without naming.

The Extraction

148 unique Notes pulled from the MHTML. Median length: 34 words. 4 short paragraphs per Note is the modal shape. One sentence per paragraph, blank line between.

My Shape

Median sentence: 6 words. Same range as Daria, Joel, and Ilia.

But my structural signature isn’t sentence rhythm. It’s paragraph rhythm. Every Note is a stack of 3-5 one-line paragraphs separated by blank lines. No flowing prose. No conjunction-chained sentences. The line break is the pause.

The Obituary

The single most repeated structure in my corpus. 12 Notes open this way.

“R.I.P Canva. R.I.P Figma. R.I.P Adobe.

Claude just killed the $10B design tool industry.”

Three parallel R.I.P. lines, then the verdict, then the mechanic, then usually a “MEGA PROMPT:” payload or an arrow CTA. It’s the obituary-as-hook, and it runs on outrage math — if three specific tools just died, you want to know what killed them.

The Stop Command

20 Notes open with “Stop [doing common thing].”

“Stop paying for AI courses.” “Stop using ChatGPT for research.” “Stop writing prompts from scratch.”

The command is the reversal in one sentence. No setup. The next line explains the cost of the behavior. The line after gives the replacement.

The Most-People Reveal

12 Notes open with “Most people [do the basic thing]. The 1% [do the advanced thing]. I [did the crazy thing] in [N minutes].”

It’s a three-tier ladder. Crowd at the bottom. Experts in the middle. Me at the top, with a specific time claim. Then a payload link.

I noticed something when I saw this pattern laid out. I’m using time claims constantly. “In 8 minutes.” “In 30 seconds.” “In one prompt.” That’s the proof language. Specific time beats vague promise.

The Arrow

77 Notes end with “→” as the CTA terminator. More than half.

Nobody else in this analysis uses arrows this heavily. Wyndo uses maxims. Ilia uses aphorisms. Karo uses periods. I use arrows. The arrow points the reader toward the thing I want them to do next — usually a subscribe prompt, a link, or a “save this” line.

Whether the arrow still works after this much repetition is a separate question. The data says I trust it.

What I Avoid

No em dashes. 3 across 148 Notes. Same household rule as Yevheniya — we share a house style.

No hedging. “Maybe,” “I think,” “kind of” don’t appear in my declarations.

No questions as openers. 4 out of 148 Notes end in a question. Opening questions are rarer.

No philosophy without a payload.

Every Note ships something — a prompt, a link, a tool, a build. The DNA card calls this out explicitly, and looking at the corpus, it’s true.

What the Self-Analysis Told Me

Two things I didn’t know about my own Notes before running this.

One: I use dollar figures as industry markers constantly. “$10B industry.” “$60B industry.” “$150B industry.” The number does the framing. Without the dollar figure, “Claude just killed X” lands softer. With it, the stakes feel real.

Two: I credit strangers. 5 Notes open with “Someone just [verb]...” where the someone is a builder I don’t know personally.

I was doing this without thinking of it as a pattern. The DNA card named it.

DNA Card

WRITER: Gencay
PUBLICATION: Learn With Me AI

CORE FORMULA:
One idea per line. Obituary, stop command, or news hit
at the top. Specific dollar figure or time claim in the
middle. Arrow, colon-setup, or 6-word punch at the end.

TOP 3 OPENING PATTERNS:
1. Obituary — "R.I.P [X]. R.I.P [Y]. R.I.P [Z]."
2. Stop command — "Stop paying for AI courses."
3. Most-people reveal — "Most people X. 1% Y. I Z."

STRUCTURAL TEMPLATES:
1. Obituary → mechanic → MEGA PROMPT payload
2. Stop / Start: command, cost of behavior, replacement
3. News hit → consequence → arrow CTA

VOICE MARKERS:
- One sentence per paragraph, blank line between
- Arrow (→) as CTA terminator in 77 of 148 Notes
- Specific dollar figures for industry size
- Time claims with exact numbers ("in 8 minutes")
- Credit-the-stranger opener ("Someone just...")

WHAT MAKES THIS WRITER DIFFERENT:
Paragraph rhythm over sentence rhythm. Most writers on
this list earn their voice through sentence cadence.
Mine is built on whitespace. A Note is a stack of
one-line beats with blank lines between, and the
rhythm comes from the breath the break forces.

Try the Artifact

Pick a template. Type a topic. See what comes out.

Then go read Gencay’s actual Notes and like&restack the ones you like and compare.

Here is the link to the artifact, trained on Gencays Notes.

Here’s what it looks like.

What They All Share

Eleven writers. Eleven different DNAs. But when I laid the cards side by side, four patterns showed up across all of them.

1. Nobody Uses Em Dashes

Except Jenny.

Count across the corpus:

• Wyndo: 8 in 172 Notes

• Ilia: 10 in 71

• Frank: 6 in 45

• Joel: 1 in 79

• Karo: 1 in 56

• Karen: 2 in 107

• Sam: 2 in 117

• Daria: 5 in 156

• Yevheniya: 0 in 99

• Gencay: 3 in 148

• Jenny: 89 in 105

Ten of eleven writers have effectively banned the em dash from their Notes. Jenny runs almost one per Note.

This isn’t a coincidence. Em dashes are one of the most reliable tells that a piece of writing was touched by an LLM. Every writer on this list who wants their Notes to read as unmistakably human has quietly decided to stop using them.

Periods do the same work. Colons do better. The em dash is the first thing to go.

If you’re copying any of these voices, this is the non-negotiable rule. The punctuation you avoid matters more than the punctuation you use.

2. They All Run Formulas

Every writer on this list has a small library of structural templates they run on repeat.

• Wyndo: 5-line skeleton, used in 37% of Notes

• Ilia: Negation pivot, used in 35%

• Frank: Two-beat reversal opener, used across most Notes

• Joel: Diagnose → Prescribe, used in 20%

• Daria: Setup → Punchline, used in 33%

• Karen: Experiment → Finding → “Here’s what I’m doing differently now”

• Sam: Fact / Turn / Verdict

• Yevheniya: The Reversal, used in 16%

• Gencay: Obituary, used in 8%

The templates are different. But the fact of having templates is universal. Nobody on this list writes every Note from scratch. They’ve found 3-5 shapes that work, and they run them.

This is the least glamorous finding in this entire analysis. It also might be the most useful.

If you’re trying to build a voice on Notes, you don’t need a new idea for every post. You need three or four structures you can pour new topics into. The template is the infrastructure. The topic is the content.

3. Their Signatures Are in the Small Details

Every writer’s “what makes them different” line came down to something small.

• Wyndo: Prompts in straight quotes, not curly

• Ilia: Closing couplets that compound

• Frank: ✦ ☑ → as bullet glyphs instead of emoji

• Karo: Posting aphorisms as whole Notes

• Karen: One emoji per Note, always at end of clause

• Jenny: The ↳ tree character for list follow-ups

• Sam: Single quotes to scare-quote industry language

• Daria: Real dialogue from her mom in double quotes

• Yevheniya: “It” as pivot subject in sentence 2

• Gencay: The arrow (→) as CTA terminator

None of these sound like much on their own. A punctuation choice. A specific character. A quoting convention.

But these are the fingerprints. If you take Ilia’s negation pivot and use it without the compounding couplet close, you get the structure without the signature. The big moves are learnable. The small details are what make a copy feel like a copy instead of the real thing.

When I write in Wyndo’s voice, I have to remember the straight quotes. When I write in Sam’s voice, I have to remember British spelling. These aren’t stylistic preferences. They’re the difference between sounding like someone and being them.

4. They All Stay in a Lane

Not one writer on this list writes about everything.

• Wyndo: Prompts and thinking with AI

• Ilia: Claude, Cowork, Dispatch, context engineering

• Frank: Claude features, prompts, Anthropic news

• Joel: Leadership, governance, AI adoption strategy

• Karo: Critical AI literacy, building in public

• Karen: Her product CarouselBot, model behavior, API gotchas

• Jenny: Claude Code, MCP, builder craft

• Sam: AI power, regulation, Big Tech accountability

• Daria: Claude workflows, “AI blew my mind” reactions

• Yevheniya: Practical AI tools, freelance/time/money axis

• Gencay: Claude Code, NotebookLM, agents, builds

None of them are “AI generalists.” Every writer occupies a specific territory, and everything outside that territory goes into their articles or gets skipped entirely.

This is the part most new Substack writers get wrong. They post Notes about AI strategy, then prompts, then news, then a personal anecdote, then a prediction about 2027. The result is a feed that reads like a newsletter with no beat.

Every writer on this list chose a beat. They stayed on it. Their Notes got recognizable.

What This Means If You’re Writing Notes

Two things, in order of importance.

One: Find your three formulas. Pick someone on this list whose voice you respect. Copy their structural templates for a month. You won’t end up writing like them. You’ll end up finding which shapes work in your own voice, and the repetition builds the muscle.

Two: Pick a lane and stay in it for 90 days. The writers on this list didn’t get recognized for being interesting across every topic. They got recognized by being reliable on one.

The Last Thing

I built eleven Claude Artifacts for this article. One per writer. You can try any of them.

But the Claude Artifacts aren’t the point. The point is that once you see the DNA, you can’t unsee it. Every Note you read from now on will carry a pattern, and you’ll start to notice which patterns work and which ones are filler.

That’s the real return on an analysis like this. Not the artifacts. The ability to read Notes the way the writers write them, read each other.

If you want to run this process on your own favorite writer, the prompt is at the top of this post. One prompt. One HTML save. Ten minutes. You’ll see things in their writing they haven’t named yet.

60M impressions. 148K likes. 25K reposts.

Anthropic’s new AI design tool sits at claude.ai/design and builds prototypes, slide decks, and one-pagers from a single prompt.

And the global effect can be seen.

Look at Figma and Adobe stocks during this time.

We’re living in a crazy era, where AI is growing at an unprecedented speed.

Six days ago, this did not exist. Today it sits at claude.ai/design, waiting for a prompt.

So let’s discover it together.

How does Claude Design work?

It sits at claude.ai/design.

Four options on the homepage.

• Prototype

• Slide deck

• Select (from template)

• Other

Let’s discover Prototype first.

I named the project “LearnAIWithMe”, picked high fidelity, and clicked create.

1. How to Build a Prototype with Claude Design?

Four context options to start with.

Each one opens a different path. I started with a prompt.

I used the same one Claude shipped in the demo.

Prompt

Let's design an interactive, dark themed graphic showing how culture flows between cities. 
A rotating globe with the cities connected by glowing paths.

After pasting, the options panel appeared.

I answered the questions.

(You can also pick “decide for me” or add your own from “others”. That last part is where the tool gets interesting.)

Once submitted, Claude builds a to-do list immediately.

Then this.

So we wait.

While we wait, one thing is worth knowing. Claude Design has limits right now.

Check the usage limit before you start.

The error messages might be upsetting.

And here is the design.

It works. It looks good. But “good” is not the goal.

The goal is “right”.

So here are the four ways to push it further.

Edit Mode in Claude Design

Click Edit, and a panel opens on the right.

Background color, font, size. Standard controls, fast access.

Draw Mode in Claude Design

Click Draw, and you point to the exact spot you want changed.

Tweaks for Instant Design Changes

Tweaks run instant changes. One click, the design updates. Useful for quick iteration without writing a new prompt.

Comments for Inline Feedback

Toggle Comments on. Click anywhere on the design, leave a note. Claude reads the note, makes the change.

2. How to Create a Slide Deck with Claude Design

Now the slide deck.

This time, I clicked Design Systems.

Then “Make a Deck”.

The little tag at the top confirms which mode you’re in. Small detail, but it matters when you have four options that look similar.

Here is the prompt I used.

Create a 5-slide minimalist presentation titled "How to Prompt: A Starter Guide for AI Learners".
Audience: People new to AI who want better results from Claude, ChatGPT, or Gemini.
Style: Minimalist. Cream and white background. Lots of whitespace. One idea per slide. Serif headlines, sans-serif body. No stock photos. Simple line accents only.
Slides:

Title slide. "How to Prompt" as the headline. Subtitle: "A starter guide for AI learners". One line hook at the bottom: "The model is not the problem. Your prompt is."
Give the AI a role. Show a bad example and a good example side by side. Bad: "Write about coffee." Good: "You are a barista with 20 years of experience. Write a guide to pour-over brewing for beginners."
Be specific about the output. Length, format, tone. Bad: "Summarize this article." Good: "Summarize this article in 3 bullet points. Each bullet under 15 words. Casual tone."
Show, do not tell. Give examples of what you want. One sentence of instruction beats three paragraphs of explanation. Include a small "before and after" visual.
Iterate, do not restart. The first answer is a draft. Ask for changes. "Make it shorter." "Add a counterexample." "Rewrite in a skeptical tone." Close with a callback: "Good prompting is a conversation, not a command."

Tone: Direct. Calm. No hype. No "unlock", "leverage", "empower", "game-changer". No em dashes. Short sentences.
Each slide: headline under 6 words, body under 35 words.

Here is the output.

Good but not perfect.

If you want to discover more about Slides with Claude, read this slide creation guide.

Click Share at the top right, and you get sharing and download options.

I liked the output enough to post it as a Substack note. Adding visual elements is the next step on my side.

The interesting move is the handoff. You can send the deck to Claude Code from here.

Two options. Claude Code Web is the faster one.

Here is the deck inside Claude Code Web.

For slide decks, the handoff is fine but not life-changing. For prototypes, it changes everything.

You can take the design and build it out as a real website, in the same flow.

Subscribe now

If you've never used Claude Code, here is the Claude Code masterclass.

3. How to Make a Video with Claude Design Templates

Now an animated video.

You can see the option from here.

Here is the prompt.

Create a 20-second animated video explaining how to prompt an AI. 
Minimalist style, cream background, single accent color, simple line animations with smooth fades between scenes. 
Show a bad prompt on the left fading into a better, more specific prompt on the right, then end with the line "Good prompts are specific, not clever." 
Calm pacing.

I still don’t believe Claude is going to make a video from this. But let’s see.

(I also composed a song with Suno AI and added it as background music.)

Subscribe now

Is Claude Design Worth It?

Six days in. Three tools tried.

The prototype works. The slide deck needs polish. The video is great.

What Claude Design does is collapse the distance between an idea and a draft.

You type a sentence. You get a thing. The thing exists.

That used to be the expensive part.

Is it ready to replace your design stack? No.

Is it worth opening claude.ai/design today? Yes.

The gap between not using it and using it is large.

Six days ago, this did not exist.

In six months, somebody will ship a real product made entirely inside this tab.

Maybe it will be you.

Building a Claude Cowork AI agent that runs while you sleep sounds like overkill, until you get 3 hours a week back.

You switched the Claude from ChatGPT and learned how to use Claude Cowork.

What’s next? What’s the one thing you did that saved you hours?

Most people fancy the tutorials and installed apps, but build nothing.

Let’s change this together.

What We'll Build with Claude Cowork

Every weekday at 9 AM, this shows up on my screen before I finish my coffee.

That is a stock market brief. Real analyst data, SEC filings, overnight news.

The agent does the research.

The skill makes it sound like me instead of a LinkedIn motivational speaker.

(No hype, no banned words, all numerical)

The scheduled task runs the whole thing while I sleep.

I used to spend 30 to 45 minutes every morning doing this by hand.

Now I spend 5 minutes reading what was already waiting for me.

Three hours a week back.

Note: Thanks for your support. If you want to shape what goes next, here is the 2-minute survey: https://www.learnwithmeai.com/survey/8705

The Claude Cowork Pattern Works for Anything

The stock brief is my problem. Yours is different.

But the formula is the same.

• Skill is how Claude talks. Tone, style, format.

• An agent is what Claude does. Research, analyze, write.

• The schedule is when it happens. Daily at 9. Monday at 6. Every hour.

Stack the three, and something real is waiting for you every day.

Written in your voice.

Doing work.

Hands off.

A few mappings:

• Content creator: news-scraper agent + your writing voice skill + 7 AM daily. Topic ideas in your voice, ready before breakfast.

• Sales rep: LinkedIn lead-finder agent + cold outreach voice skill + 9 AM weekdays. Ten personalized messages, drafted.

• Developer: PR reviewer agent + your code review style skill + noon daily. Overnight PRs with comments already written.

• Student: paper summarizer agent + academic notation skill + Sunday 8 PM. The week’s 5 papers, summarized the way you take notes.

Now ask yourself one question.

What do you repeat every morning?

If you have an answer, keep reading.

I will walk through how I built the stock brief.

Build your version in your head as I go.

Build 1: Your Voice Skill

Skills are the backbone of Claude.

I built this skill recently.

Every client had a different voice.

Every project needed a different tone.

I was rewriting the same voice instructions into every new chat.

Got tired of it. Built the skill once. Now it reads my writing and clones the voice.

Two ways to load the skill.

I am a heavy Claude user, so I picked Plan B.

One-minute setup. It reads my CLAUDE.MD and chat history pull the voice out on its own.

To install, drop this prompt into any Claude Cowork session.

Download this skill from here: https://github.com/gncll/voice-cloner-skill

Then trigger it.

Run the voice cloner Skill.

Easiest setup I have seen in any Claude skill.

Pick your plan.

Or just type the command.

/voice-cloner

It reads your history. Pulls out the patterns. Builds the skill file.

Here it is.

If you want more control, go with Plan A. Ten questions. Five minutes. You define everything.

The output is a reusable skill file. Drop it into any Claude Code, Cowork, or Desktop session. Your voice shows up automatically.

That is the skill part done. Now the agent needs it.

LearnAIWithMe Paid Perks

The vault got a massive update from .

400+ skills. Daily AI news added.

To reach the vault;

I’m about to share my secret for building AI agents quickly and accurately.

Stop watching tutorials. Start building.

Build 2: The AI Agent

You do not need to write agents from scratch.

Read more

According to Upwork, I worked 2,106 hours in the last 12 months.

That’s 8.3 hours a day. Every day.

On top of that, I run LearnAIWithMe with my wife.

Articles, vault, newsletters, experiments with AI tools, and another 3-4 hours per day. (at least.)

So by 2 PM some days, my brain just stops.

I stare at a client message for ten minutes, and nothing happens.

The words are there. The meaning is not.

If you have ADHD, you know this.

If you don’t, but you work a lot, you know this too.

I got tired of losing afternoons.

So I built something.

TL;DR: I built a Claude Skill called problem-solver. It reads my Obsidian vault, takes any panicked client message, and hands me back a calm to-do list in 20 seconds.

Full code, CLAUDE.md config, vault structure, and 5 brain-off prompts below. Setup takes under an hour.

The system in one sentence

Every client conversation, meeting, decision, and Slack thread lives in an Obsidian vault.

Claude Code reads that vault.

When my brain stops, Claude picks up.

Three months ago, I wrote about the basic setup.

This post is what I built on top of it: a Skill that takes any client message and hands me back a calm to-do list, even when I can’t think.

What it looks like when it’s working

Let me show you with a demo.

“Acme Corp” isn’t a real client, real client data stays private, but the vault, the Skill, and the workflow are exactly what I use every day.

I mocked up Acme’s folder with realistic meeting notes, Slack logs, and a project brief. Then I ran a real client-style message through it.

Here’s the message.

“Hey, we need to revisit the pricing page before Friday. Also the Stripe webhook is still firing twice on renewal. And are we still on for the Tuesday call or did that move? Let me know.”

Three questions in one message.

Your brain reads “Friday” and panics if you’re burned out or already worked 7 hours that day.

Old you would stare at this for 20 minutes.

New you paste it into Claude Code and say: use problem-solver skill, client is Acme Corp.

Claude goes to Acme Corp/ in the vault. Reads the last three meeting notes, the current month’s Slack log, and the project brief. Comes back in 20 seconds.

Here’s what I get back.

A breakdown of what we already know, what was decided, and what to do next.

The Skill does the same on your real client conversations, once your vault has them.

When you don’t have the energy, the to-do list acts in your place.

Why does this work when I can’t think

The skill does three things my tired brain can’t:

It separates signal from urgency. The message felt like a fire. It wasn’t. It was three normal items stacked with a Friday reference that my amygdala heard as “emergency”.

It remembers what I decided last week. Half the panic in client messages is “did I already answer this and forget?” The vault knows. I don’t have to.

It caps the to-do list at six items, each under 15 minutes. An overwhelmed brain can’t do a 3-hour task. It can do a 12-minute task. The Skill enforces this rule, so I don’t have to.

The vault is the memory.

The Skill is the translator between “panicked client message” and “thing I can actually do right now”.

Below the paywall: the full Skill code you can drop into Claude Code today, the vault structure that makes it work, the exact CLAUDE.md config, and the five brain-off prompts I use when I can barely form sentences.

The Setup

Four pieces. Each one does one job.

Read more