Okay, I don’t know about you, but I get confused AF about which Claude model to use when:

• How are all models different

• Which mode to use within each model

• When to use extended thinking/deep reasoning, and

• How to optimise tokens, because Claude is a bi**h when it comes to token consumption

If you’re in the same boat as me, today’s issue is for you.

I spent the last 2 days prompting each model with the exact same prompt (a simple writing task) to test how tokens are consumed. I tested the token consumption for each model with and without the extended thinking mode on, and the results were very interesting.

We’ll talk about that today.

If you like what you’re reading, subscribe to this newsletter — every other day, you’ll understand AI well enough to be the smartest person in the room at work.

Each issue is just 5 minutes — less than the time you spend doomscrolling before bed. Except, this actually moves your career forward. Join 8,000+ subscribers now!

Subscribe now

First, how does Claude’s usage work?

(If you know about tokens, skip this section)

For example

• The word “chat” is one token

• “Chatting” would break into two tokens: “chat” + “ting”

• A punctuation mark like a full stop (.) is also a token

Every interaction with the AI consumes tokens. And you get a finite number of tokens to use on your plan (free/pro/max or whatever new plan they have)

Now, Claude doesn’t tell you how many tokens you get to use as an absolute number. It just gives you a percentage of the tokens you’ve consumed. And that too is in two parts:

1. Session tokens: The minute you send the first chat to Claude, a 5-hour session begins. You have a specific number of tokens that you can use in the next 5 hours only. Once that limit is hit, your usage stops, and then you can only resume after the 5-hour window is completed. And no, Claude doesn’t tell you how many tokens you can use. It just tells you what percentage of tokens you’ve consumed. Yeah, it’s weird.

2. Weekly tokens: Claude gives you a limit for the full week. The minute you exhaust the tokens for the week, your usage stops. It gets reset only after the week is over. Again, you don’t know the absolute number of tokens you can use.

To check your usage, go to Claude.ai → Click on your name in the bottom left → Click on “Settings” → Click on “Usage”

You’ll see something like this:

How many tokens are consumed for each prompt?

Like I said, you won’t know how many tokens are consumed for each prompt/task. But this newsletter will tell you how to determine an approximate. Read on to know how..

Tokens are essentially consumed for 3 activities:

1. Input tokens: consumed when you type words and hit send (to process the input)

2. Output tokens: consumed when the AI gives you the output

3. Reasoning tokens: if you use the “reasoning” or “thinking” mode (more on that later) → uses tokens

But there are 6 models of Claude (Haiku, Sonnet 4.6, Opus 4.6, Opus 4.7, Opus 4.8 and Fable 5). Within each of them, there are different effort levels (low, medium, high, extra, max), and each of them use a different amount of tokens. And each of these can be used with or without extended thinking (deeper research).

Phew! Just typing all of this is overwhelming 😰

So there are 60 different modes that will give you different ouput for the same prompt (6 models x 5 effort levels within each x 2 (extended research)).

I basically tested each of those modes and noted the output vs consumption (reward vs cost) for them.

Now, to understand token usage between models, here’s what I did:

I gave Claude one task, to write a LinkedIn post. Gave the exact same prompt to all models and noted the usage. Here’s the prompt:

The token consumption results were really interesting. But I can’t give you a table of 60 rows, or you’ll go crazy. So I’ll talk about Claude’s default mode for each model.

Now, before you go any further, I’ll ask you to make one assumption: Instead of the token usage as a percentage that Claude gives, we’ll assume that we get 100 tokens in a 5-hour session. So just for simplicity, we’ll assume that the usage is not in terms of percentage but in terms of number of tokens. It’ll be easier to compare.

Here’s the token consumption for each model:

Reiterating: The token consumption is actually a percentage. We’ve assumed an absolute value of 100 tokens per session for simplicity, so that you can compare models easily.

There were some interesting data points and insights that popped up. Let’s talk about token consumption first.

Base mode for all models uses similar number of tokens

If you like what you’re reading, subscribe to this newsletter — every other day, you’ll understand AI well enough to be the smartest person in the room at work.

Each issue is just 5 minutes — less than the time you spend doomscrolling before bed. Except, this actually moves your career forward. Join 8,000+ subscribers now!

Whether you use Sonnet 4.6 or Opus 4.8, if you’re not using the “extended thinking” mode, it consumed anywhere between 1 and 3 tokens for a writing task.

So if you need better output (and don’t want deep thinking/reasoning), use the higher model. The cost won’t matter a lot.

Extended thinking costs 3-7x more

Well, nothing in life comes free. If you want extended thinking/reasoning, it uses 3 to 7 times more tokens than the normal mode. So use it only when needed. And make sure you check whether extended thinking is on or off before you hit enter. You wouldn’t want to waste 7x the tokens for just writing an email 🤷🏻

I’ve mentioned about where to use extended thinking and where not to use it at the end of the newsletter. Read on..

Opus 4.6 is a token guzzler

I rechecked this twice — Opus 4.6 used more or less the same tokens as Opus 4.8. In fact, it used more tokens than Opus 4.7 (the better model).

So I’m going to stop using Opus 4.6 from now on. Why pay the same (or even higher) for a worse result? 🤷🏻

Which model to use when?

Now, the interesting part.

Like I said, I gave the same prompt to all models (and sub-modes within them), and here’s what I saw:

1. Haiku and Sonnet were pretty sad for writing content. Even with extended thinking. They had bad hooks, lots of em dashes, and AI language written all over it (even after I specifically told it to avoid AI patterns, with examples). I’d never use either of those models for content generation.

   Use it for: Basic fact-finding, drafting a message, correcting grammar, converting text to table, etc.

2. Opus 4.6 onwards, the output got better. Still not usable, but at least it got the basics right. I might use it for low-effort, low-impact day-to-day stuff.

   Use it for: Writing emails, basic analysis of data

3. Now, Extended Thinking: This is the most important. The output was significantly better with extended thinking mode on. And incrementally better for each model. It did research and used data to substantiate the claims it made. Much more convincing.

   But the output was still not usable as-is. I’d need to double-check the sources, rewrite the hook and make edits to the flow.

   Use it for: Getting first drafts for LinkedIn posts or marketing-related writing that needs to convince someone. But just the first draft.

Frankly, none of the models gave me a direct, deployable output. Which is why I don’t use AI to write my posts. Sure, I might use it to brainstorm and get ideas on the writing, or at max generate a first draft. But that’s about it. It’s still not there yet. Not for something that’s a primary task for me.

Another observation I had was that without extended thinking, the ouputs for all Opus models were more or less the same. Just different language, not a different structure. Basically substantiates my earlier point about all models consuming more or less the same tokens when used without extended thinking — the ouput is also more or less similar.

So how do you use this in your daily tasks?

Well, now that you know the token consumption and the output each model gives, you can take better decisions about which model (and mode) to use in your tasks.

I’d say for any of your important tasks, you would need extended thinking. It’s the bare minimum. Yes, it’s expensive, but it eventually saves you time if you do it right.

(I’ve written about how to conserve tokens in a separate post. Check that one out too)

Do note that while I’ve used WRITING as a task for Claude in this example, the foundation still stands — Haiku and Sonnet are still very basic; and Extended Thinking is still something you must use for any important task — whether you’re writing a post, creating a presentation or doing research.

What next?

I haven’t covered the research mode yet in this newsletter. I was literally out of tokens, and wondering if Claude will think I’m an abusive user 😂

So if you want to read about the “Research” mode, comment on this post or reply to this email, and I’ll write about it next week 😊

Well folks, that’s all for today! If you enjoyed reading this one, share it in your network. Every share helps get more subscribers.

I’ll see you next week!

Cheers,

Ankur

The thing is, I’ve seen that most AI content is either clickbait, or superficial, or just too technical — meant for engineers. So I’m building The Decodr — a platform where non-tech professionals come to understand and use AI. Deeply researched stuff, free of the noise.

It hosts a newsletter with deep researched content on AI — explainers, how-to guides, platform comparisons and more — all meant for non tech folks. But that’s not it.

On top of that, we’re also building repositories of ready Claude skills that you can use to build stuff. We’ll compare tools, tell you what’s best for your use case, and also build a community slowly.

So before we go all-in, I’d like to know if you’d even use a platform like that. Check out the website and (PLEASE) tell me if you’d use it. It’ll help me understand what users want :)

Now, here’s what today’s issue is about:

And the proof is actually in the pudding this time: I

- purchased a domain
- built TheDecodr with Claude
- hosted it on Github (no, I’m not a coder; did it for the first time), and…
- got everything up and running, from zero, in under 6 hours!

And it’s the first time I did something like this — was genuinely proud of myself (and Claude, obviously), so I thought “Damn! THIS is what the next LazyAI edition should be about!”

So here we are.

Today’s read might be a bit long, but if you read it fully, I promise you can build what I built in less than half a day. If you don’t have time today, star mark this email and come back to it later.

But please do it. Just experience the joy of building from scratch once. It just hits different!

Let’s begin.

First, get the domain

The domain is the name of the website (www.yourbusiness.com). Whatever project you’re working on, think of a name you’d want to call it, go to a website like GoDaddy, BigRock or any other, and purchase the domain. The exact same name may or may not be available, so you may need to look at combinations.

Below is how you search for a domain.

Buy the domain you like. It’ll cost you about INR 1,000 a year.

If you don’t want to invest in the domain right now, that’s cool too. You can just use a free hosting service and still have your app up and running (more on that later). Although I’d advise you buy to it; because making a monetary investment is one of the best ways to get consistent at something :)

Moving on..

Now, start building with Claude

This is the meat of the entire operation. Open Claude, and just talk to it like you would talk to a product manager who is helping you make an app/website. In English language.

Step 1: Brainstorm with it and create a PRD

Start with the prompt:

I want to build a website/web app about [My project]. Can you help me run through how it should be built. I have some ideas, but I want to brainstorm with you and generate a PRD at the end of it.

If you’re making a basic app, make sure you use “Haiku 4.5” as the model. If you don’t like the output you get after a few iterations, then use Sonnet 4.6 — Medium.

Do not use Opus as the model. It will burn your usage within minutes!

Now, continue talking with Claude until all your specifications are clear. It’ll tell you how your website will be made. Then you ask it to create a PRD (Product Requirements Document) for it and download the .md (markdown) file of the PRD.

Step 2: Build it

Start a new chat. In the new chat, upload the PRD and ask Claude to build the app/website. Tell it that you’re not a technical person, so in any chat going forward, it should not use jargon.

The reason I’m asking you to create a PRD and start a new chat is because it’ll save tokens. Continuing a chat with 50 messages will hit your usage limit in minutes.

Claude will give you the actual website/app with the UI.

Step 3: Iterate

Once you get the output, iterate on it. Tell Claude things like “The button colour is not appealing. Change it. Use design pricniples as a designer would”

With a few iterations, you’ll get a nice-looking app/website.

Now, let’s make it live!

There are two ways to make your website live (in technical terms, deploy it):

1. Directly on the hosting platform (more on that later), or

2. Via Github (more on that later too)

You can use either, but I’d say use Github — it’s a cleaner way to do it, and also keeps track of all the changes you’ve made if and when you decide to update the app.

(And it’ll also give you developer feels when you understand and use words like Git, repo, commit etc. 😉)

Step 1: Sign up on Github and create your repository

Go to Github.com and register for a free account. Create a new folder (called as repository/repo). This is where you’ll be storing all the files of the project. The files work in tandem with each other (Claude would have already ensured that).

Use the below settings while creating the repo:

If you’re stuck somewhere, take a screenshot and ask Perplexity how to proceed.

I’m asking you to use Perplexity because it’s more accurate for things like this, and because Claude burns a lot of tokens for tasks. So don’t waste it on troubleshooting.

Step 2: Export the files from Claude

Ask Claude to give you all the files of your website/app. Download all of them in your computer. It might give you a zip file or not, depending on the project (and its mood :p).

If it’s a zip file, unzip it on your computer.

Step 3: Upload files on Github

Go to your repo, click on “Add file” and click on “Upload files” and drag-and-drop the files from your local system into Github.

Again, if you’re stuck somewhere, take a screenshot and ask Perplexity how to fix it.

Now here’s some nuance, just for your understanding:

Typically, the index.html file is the default home page or main entry point to your website. When a user types a website address (www.yourwebsite.com) in the browser, the web server automatically looks for this index file and displays it as the default landing page. Then this file internally navigates the user to different pages depending on how the user scrolls through the website and where he clicks.

Anyway, now the files are uploaded to Github. But your website isn’t live yet.

Step 4: Hosting your website

Now, you need a place on the server where you will store your files. That’s the hosting provider.

Think of it this way:

• Domain: It’s just the address of your website (like an address to your house)

• Hosting provider: This is actually your house where your website files live and are available to see

• Github: This is like a house construction notebook (or a blueprint) where you write down every change to your house design, and once approved, copy the changes on the actual house where your files are hosted

You can use any hosting provider that gives a free plan. I used Netlify.

So you upload the files on Github and connect it to Netlify. Just go to netlify.com and click on “Add New Project”. It will then give you the option to connect with your Github repository. Fairly straightforward.

Once you do that, choose the free hosting and deploy the website.

That’s it. It’s live!

Now, whenever you make subsequent changes to your project on Github and COMMIT the changes (which basically means approve the changes), it will automatically sync with Netlify and update your website within minutes.

But the story doesn’t end here. If you’ve purchased a domain, you need to connect your domain to Netlify’s hosting address.

I won’t get into the details of that, or this article just won’t end. You can do a quick Google search. It’s pretty straightforward.

But to begin with, just try the free hosting without the domain. Netlify will give you a free sub-domain (something like your-app-name.netlify.com). Use it and see how it feels. If you want to pursue it seriously, get the domain, If it’s just a fun project, use the free hosting and sub-domain with Netlify.

Now, what if you want to make changes to the website in future?

There are 2 ways of doing it:

1. The crude way: Interact with Claude in a new chat and give it context everytime.

2. The better way: Create a project on Claude with all the relevant instructions fed to it one-time, and just make changes within that project. whenever needed I’ll quickly touch up on how to do that.

For future updates, create a Claude project

The same chat window where you made your project by talking to Claude and exported your project files? You need to tell Claude to do one last thing. Give it this prompt.

“I want to create a Claude project for this website. Give me a detailed instruction file that will have all relevant information and context about this project: Objective, branding, messaging, and anything else that you feel is important for Claude to understand the complete context. I will upload this as the instructions in the Claude project, so I want you to be detailed such that you don’t lose any context. But don’t be verbose. Try to keep the length of the instruction file optimum.”

That’s it. Claude will generate an instruction file for you. Copy the content.

Now go to “Projects in Claude → Create a new project → Give it a name and description, and you’ll see the below screen:

Leave the Memory section alone. Claude will auto-update it every evening based on what you do with the project.

Open the Instruction section and paste the instruction file that Claude gave you in the earlier step. This will give Claude all the context it needs for updating your app/website in future.

In the files section, upload the index.html file and any other files that you feel are important. In fact, you can also sync your file with Github — just click on the “+” sign and click on Github. It’ll ask you for your Github credentials and connect it automatically.

Make sure you don’t upload all files in this section, since it has limited memory. Only the important ones.

Now, how do you make future updates?

Whenever you want to change anything, just go to the Project and start a new chat there. Talk to it like you’d talk to a product manager/engineer building this.

It’ll have all previous context and files, so it can make all the required changes.

Once you’re happy with the changes, ask it for the updated file, and upload it to Github the same way you uploaded the files earlier.

That’s it. Your website will now reflect the updates.

And that’s how you make your passion project come to life. With Claude.

If you read it till here, please check The Decodr and tell me what you think. It’s going to have a lot more content like this, along with tools, Claude skills, and more — for non-technical people who want to understand and use AI in their work.

If you want to talk about this or anything else, please reply to this email or drop it in the comments. I would love to read what you guys have to say!

I’ll see you next week!

Cheers,

Ankur

- “AI just killed copywriters”
- “Analysts are dead”
- “Claude is my content team”

But here’s a twist: Anthropic, the company that buit Claude, is hiring for

- a HUMAN copywriter
- a HUMAN accounting manager
- a HUMAN content head

For salaries in the range of $190K - $400K.

Read that again: Anthropic is paying up to $400K. To a human. For doing a job that people are saying it’s own tool has replaced.

What in the world is happening?

No, it’s not hypocrisy. Anthropic just knows what Claude is capable of, and where it needs humans.

And that’s what I’m going to be talking about in today’s newsletter. Tell me if you’d like more pieces on such topics, I’ll incorporate them in future editions :)

You don’t need to be an engineer to talk AI. Subscribe to this newsletter — every other day, you’ll understand AI well enough to be the smartest person in the room at work.

Each issue is just 5 minutes — less than the time you spend doomscrolling before bed. Except, this actually moves your career forward. Join 8,000+ subscribers now.

Subscribe now

There are a few problems with AI today that are preventing it from replacing humans.

Problem 1: AI doesn’t have taste

I had written about it a few weeks ago. AI is based on probability. So it will generate an output based on patterns, not taste.

Basically, if you ask anything to an AI, it will give you the most “probable” answer. The one that has the highest probability.

This means that the output it gives you, will be more or less similar in depth to the ouput it gives your competitor. Or your colleague who’s 5 years junior to you. Or a second year college student. Which in turn means that you either accept that your work is equal to that of a college student’s, or you give it deeper instructions, a better prompt.

And that better prompt? That comes with taste. From knowing what doesn’t work after years of experience. From knowing that your user might not like what is produced. From knowing things that are not documented on the internet. Things that you just..know.

So that AI needs taste — which means it needs a human with expertise to guide it. Which eventually means that it cannot replace one.

Problem 2: AI is not accurate

You can never be 100% sure of AI output. And the AA-Omniscience benchmark shows why.

AA-Omniscience is a benchmark that tests how well AI models know facts, and whether they admit uncertainty instead of guessing.

Now, this study involved asking each AI model 6,000 questions across 6 domains to see if the AI

• answered them correctly

• said that it doesn’t know the answer, or

• just hallucinated to give some garbage

(The caveat was that these were difficult questions which only experts in their field could answer)

Below were the results. They’re surprising.

The highest accuracy was 57% for GPT 5.5 max.

Claude Opus, which we know (or rather believe to be) as the smartest model out there, answered only 46% of questions correctly.

Point being, AI can’t be trusted yet. Not for difficult questions/tasks at least. You still need a human for that.

Problem 3: The people maze

You don’t need to be an engineer to talk AI. Subscribe to this newsletter — every other day, you’ll understand AI well enough to be the smartest person in the room at work.

Each issue is just 5 minutes — less than the time you spend doomscrolling before bed. Except, this actually moves your career forward. Join 8,000+ subscribers now.

Subscribe now

At the end of the day, an AI is still a computer. Even if it might have good hard skills in future, it has zero soft skills. And that’s where humans have the edge.

When you’re planning a project with multiple stakeholders, the technicalities are the easier problem to deal with. The more difficult problem to deal with, is people.

Stakeholder management, team management, upward reporting — navigating the people maze, is what AI sucks at. And that’s THE most important thing in a corporate job.

In fact, Daniela Amodei, the co-founder of Anthropic, explicitly mentioned that they “look for people who are great communicators, who have excellent EQ and people skills, who are kind and compassionate and curious and want to help other people”

So even if AI builds taste and does tasks with 100% accuracy (sometime in the future), soft skills and stakeholder management are not something it will be able to do. Not today. Not 10 years later.

And without that, a company cannot function.

So what does it mean for you?

If you’re an Individual Contributor, there are 3 things I’d want to leave you with:

1. Build expertise and taste. Get deep enough in your function that you can question AI’s output. Marketers should be able to challenge a weak hook or a story with no tension. Finance folks should understand when a model’s assumptions are off before checking the formula. Product folks must know when a feature solves the wrong problem. Build that depth so that you’re able to question AI

2. Polish your communication and people skills. AI handles the output. You still have to win the room. Irrespective of your function, know your stakeholders well — clients, users, cross functional teams — anyone. That’s a gap AI will never be able to fill.

3. Verify every claim AI makes. Verify all claims that AI makes. Please. I cannot stress this enough. You saw the accuracy of AI. So just make sure that you’re verifying every single claim that your AI tool makes when you interact with it.
   Because what’s worse than getting replaced by AI, is getting replaced by a human — because you goofed up with your data.

If you’re a manager/leader:

1. That expertise we spoke about? You need to build it too. Not because you need to use AI on a daily basis. But because you should be able to tell your team when to use it and when not to use it.

2. Communication is even more important for you. Your team will move faster with AI. Your job becomes less about output management and more about direction-setting and stakeholder alignment. That's a harder skill, not an easier one.

3. Lastly, and most importantly, do not expect your team to give you 10 hours of worth of work in 1 hour, just becaue “AI can do it with one prompt”. That’s the 2024 narrative. Unless you want AI slop, that won’t happen.

AI is not there yet. Stop expecting it to be.

Yes, the world is getting tougher. Entry-level jobs are going. But if you’re a mid-career professional, you’ll see the hype calm down in the next one year.

Till then, just hold on tight.

That’s it for today folks! If you found some solace after readng this, please share it with a friend. Tell them to breathe. It’s okay :)

Cheers,

Ankur

If you’ve ever stared at a “you’ve reached your usage limit” screen on Claude and wondered what you did wrong, this article is for you.

It’s not the questions. It’s the conversation.

Picture this. You open Claude on a Monday morning. You ask it to help draft an email. Then you tweak it. Then you ask a follow-up. Then another. By noon, you’ve had a 25-message back-and-forth — and suddenly, Claude tells you you’re out of credits for the day.

What happened?

Well, here’s what you should know: Claude doesn’t just read your latest message. It re-reads the entire conversation every single time you hit send. Right from message #1.

For every conversation, it uses tokens.

A token is simply a chunk of text. AI doesn’t read sentences the way we humans do. It breaks down sentences into tokens, each of which could be a whole word, part of a word, or even a punctuation

(I’ve written more about tokens in an earlier newsletter. You can read about them below:

Now, Chatbots consume tokens for every single chat message in the window. And that’s why a 25-message conversation feels like it eats up your credits. Because it’s not 25 separate questions. It’s 25 questions where each one carries the full weight of everything that came before it.

One creator tracked his Claude usage in detail and found that 98.5% of his tokens were going towards re-reading conversation history. Not the actual work he was asking Claude to do.

Read that again.

You don’t need to be an engineer to talk AI. Subscribe to this newsletter — every other day, you’ll understand AI well enough to be the smartest person in the room at work.

Each issue is just 5 minutes — less than the time you spend doomscrolling before bed. Except, this actually moves your career forward. Join 8,000+ subscribers now.

So how do you ensure you don’t hit Claude’s limits?

Once you understand the token problem, the fixes are simple. Every habit below is answering one question: How do I reduce unnecessary token consumption?

1. Start a new chat for a new topic

This is the one people resist most, because starting a fresh chat feels like starting over. It’s not.

When you switch topics inside the same conversation, you’re forcing Claude to carry the full history of the previous topic into every reply about the new one. Imagine asking your accountant a tax question in the middle of a meeting about your marketing strategy, and they insist on re-reading the entire marketing discussion before answering. That’s actually what’s happening.

New topic. New chat. The context you actually need takes 10 seconds to re-paste. Everything else is just weight.

2. Edit the message, don’t send a correction.

When Claude doesn’t get it right, most people just type “make it shorter” or “no that's not what I want. I want XYZ” and hit send.

But what that does is, it adds one more message to the stack that Claude now keeps re-reading forever.

(Before you scroll ahead, I’m going to use Claude in this article, because this article IS about Claude. But the structure and process is pretty much the same across all Chatbots. ChatGPT, Gemini and the rest — all follow the same token usage process)

In Claude’s chat interface, you can click Edit on any message you’ve already sent, change what you wrote, and regenerate the response. The new version replaces the old one. It doesn’t stack on top of it.

This single habit probably saves more tokens than anything else on this list. Please use it. Aggressively.

If you like what you’re reading, subscribe to this newsletter — every other day, you’ll understand AI well enough to be the smartest person in the room at work.

Each issue is just 5 minutes — less than the time you spend doomscrolling before bed. Except, this actually moves your career forward. Join 8,000+ subscribers now!

3. Ask for everything at once.

“Summarise this article.” (reads response) “Now pull out the three key points.” (reads response) “Now suggest a headline.”

That’s three separate messages. Three full context reloads. Claude re-reads the entire conversation each time.

Instead, keep it one message — “Summarise this, pull out three key points, and suggest a headline”. It’s one reload. Same output. The answers are usually sharper too, because Claude sees what you’re building toward from the start.

4. Don’t redo the whole thing when one part is wrong.

“Rewrite this email” asks Claude to regenerate every word. “Only fix the opening line — keep the rest exactly as is” asks Claude to touch one sentence.

If your email is 300 words and you ask for a full rewrite, you’ve just spent that many output tokens fixing something a 20-token surgical edit would have handled.

Be specific about what’s broken. Point to the exact paragraph. And add “no explanation needed, just the updated version”. Otherwise Claude will spend a paragraph telling you what it changed. And that paragraph costs tokens too.

5. Summarize and restart when a conversation gets long.

Every conversation has a point where it becomes more expensive to continue than to start fresh.

When you feel a session getting bloated — say, past 20-25 messages, ask Claude to write a summary of everything important so far. Copy it. Open a new chat. Paste the summary as your first message. And then continue from there.

You lose nothing. You carry forward the context that matters, without the token cost of re-reading every exchange that got you there.

If you like what you’re reading, subscribe to this newsletter — every other day, you’ll understand AI well enough to be the smartest person in the room at work.

Each issue is just 5 minutes — less than the time you spend doomscrolling before bed. Except, this actually moves your career forward. Join 8,000+ subscribers now!

6. Use Projects for files you keep going back to.

Every time you upload a document to a new chat, Claude processes it fresh. Upload the same 10-page brief to five different chats and you’ve paid to process it five times.

Claude’s Projects feature lets you upload a file once and reference it across multiple conversations without that repeated cost. It also uses smarter retrieval — pulling only the sections relevant to your question instead of loading the whole document every time.

If there’s a document you return to regularly, like a brand guide, a brief, a research report or something else — just move it into a Project. And then just use that project’s environment for new tasks.

Okay, quick poll. I want to understand if you guys have ever used Claude projects. If not, I’ll write a detailed article on it next. It’s one of the most underused features of Claude!

7. Choose your model carefully

Claude has three main models: Haiku (fast, light, inexpensive), Sonnet (balanced), and Opus (the heavy one, best for complex reasoning).

Most people leave it on Sonnet or Opus by default. And then wonder why their credits disappear quickly. Haiku handles a surprising amount: summarising text, fixing grammar, reformatting a document, brainstorming ideas. It’s not a downgrade for those tasks. It’s the right tool.

Save Opus for the work that genuinely needs it, like nuanced analysis, complex writing, and situations where the quality difference is actually visible. For everything else, Haiku is fast, capable, and uses a fraction of the cost. Or at least stick to Sonnet.

8. Stop asking Claude to do things it’s not built for.

This is the most important. We’re all lazy, and we want one tool that does everything.

Now, Claude is exceptional at reasoning, writing, and analysis. It cannot generate images. It’s not the fastest at real-time search.

If you spend six messages trying to get Claude to produce a visual — describing the colours, the layout, the style — you’re burning tokens on a task it was never designed to solve. Use Gemini. Same logic for live news or real-time data: there are better tools for that job.

I’m not saying Claude is limited. I’m saying every tool has a lane. Keeping Claude in its lane isn’t a workaround — it’s just using it well.

So what does this mean for you?

The usage limit is not arbitrary. It’s just math — and most of us are doing the math wrong without realising it.

Pick just three habits from this list. Start with: new chat per topic, edit instead of correcting, and ask for everything at once. Those three alone will feel like a noticeable difference by the end of the week.

The limit stops being a wall once you stop carrying the whole conversation on your back.

Let me know if this helped!

See you next time..

Cheers,

Ankur

AI slop is the flood of low-quality AI-generated content, like text, images, audio, and videos, often created for clickbait, ad revenue, or mass production rather than real value.

53.7% of posts written on LinkedIn in 2025 were AI. Yes, more than half of the stuff we’re reading is AI slop. No wonder that the term was added to the Cambridge dictionary as more and more internet users kept complaining about it.

It took me some time to understand what was happening. And once I did, it was clear.

What I was noticing was AI’s fingerprint. And once you understand why that fingerprint exists, you’ll spot it everywhere — LinkedIn posts, comments, Instagram Reels. In about 10 seconds.

First, why does AI have a fingerprint at all?

When an AI model like ChatGPT or Claude is trained, it learns by consuming billions of pieces of human-written text — articles, books, social media posts, forum threads, everything. It then learns to predict: "given what came before, what word or sentence is most likely to come next?"

It doesn’t actually think. It looks at what you asked, cross-references everything it was trained on, and produces the statistically most likely response. It's just completing a pattern — very, very convincingly.

Think of it like your phone's autocomplete, but on steroids. When you type "Happy" on your keyboard, your phone suggests "Birthday". Because that's what billions of people type after "Happy." It's not thinking. It's just repeating the most common pattern it's seen. AI writing works the same way, just at a scale that can produce 500 words instead of two. Ask it for a "motivational LinkedIn post" and it reaches for the most common patterns it's seen for that phrase — because statistically, that's what works. Every time.

You don’t need to be an engineer to talk AI. Subscribe to this newsletter — every other day, you’ll understand AI well enough to be the smartest person in the room at work.

Each issue is just 5 minutes — less than the time you spend doomscrolling before bed. Except, this actually moves your career forward. Join 8,000+ subscribers now.

Subscribe now

Now that you know that there are tells, let me tell you how to identify them.

On LinkedIn posts

You’ll see one of more patterns when someone has written a post with AI.

1. Lot of high-impact words: Posts will have words like “boasts a, exemplified, terrible, quietly’ etc.

2. Rule of three: This is a powerful concept in storytelling, where sentences are structured using groups of 3 to have higher psychological impact. For example, “He lost his job. His money. And the respect that came with it”.

3. Phrases that show exclusivity to the writer: Things like “Here’s what nobody’s telling you”, or “Here’s why nobody talks about it”. These phrases are used because it makes the reader feel like they’re reading exclusive information. The problem is, every second writer uses them :D

4. Antithesis in structure: Sentences like “It’s not A. It’s B” or some variant of them

5. Overuse of em dashes (—) in the post

I could go on and on. The list is endless. But these are the most common AI patterns you’ll see in LinkedIn posts.

One thing to keep in mind though: Not everyone who writes using these patterns is using AI. Some people genuinely write well. A lot of people use em dashes naturally (I’m one of those). The key is to see if it’s overdone. If it is, it’s mostly AI.

On LinkedIn comments

This is where it gets sad.

1. AI comments don’t engage. They validate. “Such a valuable perspective!” “This really resonates.” “Thanks for sharing this insight!”

   These aren’t thoughts. They’re the text equivalent of a “I agree”.

2. A real comment reacts to something specific in the post. An AI comment could be copy-pasted onto any post on any topic and still make sense. If it fits everywhere, it came from nowhere.

Interesting fact: There are “engagement pods” on LinkedIn. These are tools like Crosslike.club, where you sign up, login with your LinkedIn credentials, and become part of a community of thousands of people. Every time anyone from that community posts on LinkedIn, randomly selected profiles of real humans from that community write comments on the post All automated in the background using AI, they don’t even know it’s being done. Here’s an example I saw yesterday (the guy shouldn’t have tagged me :))

3. If an engagement pod is being used, you’ll see another pattern: Timing. AI- generated comments come in consistently, especially in the first one hour of the post going live. Because that’s what the algorithm rewards.

Essentially, if it’s all praises in the comments section of the post, it’s most likely an AI.

On Instagram reels

I had written an entire newsletter on this. There are fake AI profiles on Instagram with more than 200,000 followers. They even get “paid subscribers” and make money off of it.

The images you see above, are all AI profiles. And here’s how you can identify them.

1. Skin looks too perfect. Real skin has pores, uneven texture, asymmetry. AI skin often feels polished or “designed”

2. Lighting and backgrounds don’t fully match. Faces lit differently from the scene, or subtly changing locations across posts

3. Hands, ears, teeth still glitch. Finger counts, mismatched earrings, unnatural teeth are common tells

4. Very few videos. Most AI influencers rely on photos because motion and expressions are harder to fake convincingly

5. Digital footprint feels empty. No old tagged photos, interviews, random friend uploads, or history outside the account itself

You don’t need to be an engineer to talk AI. Subscribe to this newsletter — every other day, you’ll understand AI well enough to be the smartest person in the room at work.

Each issue is just 5 minutes — less than the time you spend doomscrolling before bed. Except, this actually moves your career forward. Join 8,000+ subscribers now.

Subscribe now

So where does it leave you?

The AI fingerprint isn't going away. If anything, it's going to get harder to spot as models get better. But AI can’t replace a real human experience. Nor can I replace real human writing.

So after reading this, you’ll know if something you read/watch is AI or not.

If you found this useful, please share it with a friend so that they don’t get fooled (or worse, scammed) by an AI.

Share

I’ll see you next time..

Cheers,

Ankur

Today, I want to talk about how I improve accuracy with my research, and why everyone MUST use Perplexity.

Your AI “remembers” more than it “thinks”

How many times have you used ChatGPT to research something — got a confident, well-structured answer — and then found out the statistic it cited doesn’t exist?

Happens to everyone.

But here’s the thing: it’s not a bug. It’s a feature of how most AI tools are built. And once you understand it, you’ll immediately know which tool to reach for — and when.

So What’s Actually Going On?

Most AI tools — ChatGPT, Claude, Gemini — are built on something called a Large Language Models (LLM).

An LLM is an AI that has been trained on an enormous amount of text from the internet — articles, books, forums, everything. It learns patterns from all of that, and when you ask it something, it generates a response based on those patterns. It’s completing — very, very sophisticatedly.

Think of it like a very well-read friend who read everything on the internet up to a certain date, and then went completely off-grid. Ask them something recent? They’ll answer confidently — based on what they last read. Whether that information is still accurate? That’s a different question.

And when they’re asked something they don’t know, they’ll just make something up.

That’s Claude, ChatGPT and Gemini for you

But Perplexity is built differently

Perplexity isn’t really a chatbot. It’s closer to a search engine with a brain.

Every time you ask it something, it goes to the web first — before generating a single word of response. It pulls relevant pages, reads them, and then builds its answer from what it just found. The language model’s job is just to translate those search results into something readable.

The internal design principle Perplexity built this around: “You are not supposed to say anything that you didn’t retrieve.”

That one sentence explains why using Perplexity feels different. It can’t invent a source. It can only cite what it actually found. And it shows you those sources, inline, next to every claim.

You don’t need to be an engineer to talk AI. Subscribe to this newsletter — once a week, you’ll understand AI well enough to be the smartest person in the room at work.

Each issue is just 5 minutes — less than the time you spend doomscrolling before bed. Except, this actually moves your career forward. Join 8,000+ subscribers now.

Subscribe now

But it’s not magic

I’ll be honest, because I don’t want to hype things just for engagement.

Perplexity still hallucinates. In one independent test, models were asked about a research paper that didn’t exist. ChatGPT, Gemini, and Claude all said they couldn’t find it. Perplexity fabricated an answer — and cited sources that didn’t support the claim.

How? When a search returns nothing useful, the model still tries to fill the gap. And when it does that, you’re back to the same hallucination risk as any other LLM.

That risk is lower with Perplexity than with other AI chatbots, but it’s still there.

Here’s the practical way to know: If a paragraph in Perplexity’s answer has no citation number next to it, the AI is speaking from training, not retrieval. That’s your cue to verify.

So what does this mean for you?

Here’s the thing: Perplexity reads the web for you. ChatGPT and Claude think for you. Both are useful, and neither replaces the other.

Use Perplexity when you need facts you can trace back to a real source — research, current events, checking if something is still true. Reach for ChatGPT or Claude when you need the AI to write, reason, or create something from scratch.

Here’s how I use it

I’ve started using Perplexity as the final step before I publish anything — articles, newsletters, social posts. I paste in whatever I’ve written and ask it to verify every claim I’ve made.

In all the time I’ve been doing this, I’d say more than 90% of the times, the sources are correct. With ChatGPT and Claude, that’s lower.

No, you will not find 100% accuracy anywhere, and manual checks will still be needed. But with Perplexity, the errors are lower, which just saves your time

Try it before your next piece of writing goes out. It takes 2 minutes, and it might be exactly what you need.

You don’t need to be an engineer to talk AI. Subscribe to this newsletter — every other day, you’ll understand AI well enough to be the smartest person in the room at work.

Each issue is just 5 minutes — less than the time you spend doomscrolling before bed. Except, this actually moves your career forward. Join 8,000+ subscribers now.

Subscribe now

Before you go, I have one request

I’ve been writing for over 3 months now, and I would love to know if I’m on the right track. So I need you to help me with two things:

1. Tell me if the weekly articles are adding value

2. Tell me what you would like to read on, so that it’s just more relevant for you

It’ll take a minute to reply, but will improve the quality of my work by 10X, so if you think my work has added ANY value to your AI learning journey, I’d love to see a reply to this email with your thoughts 😊

I’ll see you next week!

Cheers,

Ankur

Yes, the email subject seemed counter-intuitive. An AI newsletter saying DON’T keep up with AI?

But hear me out..

Today’s piece is an attempt to calm the panic that hits when you open LinkedIn and every third post is about an AI tool you’ve never used.

Open any feed for 30 seconds and you’ll see it — someone teaching prompt engineering, someone evangelising agentic workflows, someone showing off a Veo 3 clip, someone selling a Claude skills course. Every one of them is good. Every one of them is making you feel like you should be doing what they do.

So you start dabbling. A bit of this, a bit of that. Two weeks later you’ve forgotten most of it, a new tool drops, and the cycle repeats. I’ve done this. In fact most days, I’m still doing it.

But the panic is built on two assumptions that fall apart the moment you examine them.

The two assumptions

Assumption one: the things you’re missing today will still matter tomorrow.

They won’t. Most of them. Because AI tools have a brutally short half-life — most of what’s trending this month is forgotten in month twelve.

Half-life is the time it takes for something to lose half its relevance

Think about it. 18 months ago, people were paying for ChatGPT prompt packs on Gumroad. Entire newsletters were dedicated to AutoGPT. “Prompt engineer” was a real job title at high salaries.

Today? Prompt packs are obsolete. AutoGPT is a footnote. Prompt engineering as a standalone role is mostly gone. And Sora — the AI video tool everyone was obsessed with six months ago — already got shut down.

Look at what’s trending right now. Some of it will matter in 2027. Most of it won’t. You just don’t know which is which — and neither does anyone teaching it on LinkedIn (this is why everyone sounds so confident — confidence in this space is mostly performance).

Spread your 5–7 hours a week across six trending things, and you’re memorising train timetables that get rewritten every quarter.

You don’t need to be an engineer to talk AI. Subscribe to this newsletter — every other day, you’ll understand AI well enough to be the smartest person in the room at work.

Each issue is just 5 minutes — less than the time you spend doomscrolling before bed. Except, this actually moves your career forward. Join 8,000+ subscribers now.

Subscribe now

Assumption two: Each new AI skill is a separate thing you should learn from scratch.

Also wrong. And this one’s the good news.

The skills underneath AI tools are supremely transferable. Spend 100 hours building Claude skills — really learning how to scope a model’s behaviour, structure instructions, handle edge cases — and you’ve also picked up most of what you need for agentic workflows, custom GPTs, and even Claude Code. Because at the core, they’re all the same skill: Telling a model precisely what you want, in a way it can actually act on.

Same with image and video generation. Get genuinely good at one model — Midjourney, Veo 3, Nano Banana, pick your weapon — and the iteration loop you’ve internalised (prompt, judge the output, adjust, repeat) is exactly the loop you’ll use for every visual model that ships next year. The interface changes. The instinct doesn’t.

The generalist stays a beginner everywhere. The deep person becomes a fast learner everywhere. Those are very different futures.

Here’s what actually bothers me

We’ve been sold the idea that staying current with AI means staying broad. It doesn’t. Broad familiarity is what everybody already has. It’s worth nothing.

And here’s the part that should make you exhale — the depth you need is probably already sitting in front of you.

According to a study conducted last year, 66% of ChatGPT users only use it for basic stuff. They type a question, copy the answer, close the tab. They’ve never set up a project. Never written a custom instruction. Never used artifacts properly. Never chained two prompts. Never built a reusable system prompt for the work they do every single day.

The gap between a casual user and a deep user of the same tool is enormous. Way bigger than the gap between a casual user of one tool and a casual user of five.

So the highest-leverage move isn’t a new tool. It’s going embarrassingly deep into the one you already open every day.

You don’t need to be an engineer to talk AI. Subscribe to this newsletter — every other day, you’ll understand AI well enough to be the smartest person in the room at work.

Each issue is just 5 minutes — less than the time you spend doomscrolling before bed. Except, this actually moves your career forward. Join 8,000+ subscribers now.

Subscribe now

So what should you take from this?

Tonight, don’t bookmark another tutorial. Just open the AI tool you already use. Ask yourself how many of its features you’ve actually pushed against. Projects? Custom instructions? Artifacts? Memory? Whatever your tool’s version of those is.

That’s where your depth is hiding. Not in the next thing on your feed.

You’re not behind on AI. You’re behind on the AI you already have. That’s a much smaller, much kinder problem to solve.

Breathe. It’s okay.

If this got you thinking, please share it with someone who keeps bookmarking AI tutorials they’ll never finish.

It takes a lot of time to research and write this. Sharing means a lot 🙂

See you next time..

Cheers,

Ankur

Today’s piece is about the AI job panic. The one in every WhatsApp group, every coffee chat, every uncle at every wedding. “Beta, will AI take your job?”

Most people are worried about the wrong thing. And I’ll tell you why.

What everyone’s reading

Last week, Goldman Sachs put out a number. AI is wiping out about 16,000 US jobs every month. Most of them entry-level GenZs.

Then a separate poll came up. 1 in 5 American workers said AI is already doing parts of their job.

Yes, sounds scary. But there are two studies that came out last year, which are not published in the media. Because they’re..well..not dystopian.

Let’s go through them

Study 1: AI vs doctors

A research team pulled together 83 different studies that compared AI diagnoses against actual doctors.

The finding: It performed almost as well as physicians overall and non-expert physicians, but it was 15.8% worse than expert physicians

Read that again.

Not “AI is amazing.” Not “AI is useless.” Just, “AI matches the average, but loses to the experts”

You don’t need to be an engineer to talk AI. Subscribe to this newsletter — every other day, you’ll understand AI well enough to be the smartest person in the room at work.

Each issue is just 5 minutes — less than the time you spend doomscrolling before bed. Except, this actually moves your career forward. Join 8,000+ subscribers now!

Study 2: AI vs senior developers

This one’s crazier.

A research group called METR took 16 senior developers — people with 5+ years on the job, working on their own code. Real bugs. Real features. Nothing fake about the setup.

And they split the work randomly: Half the tasks with AI, and half without. Then they measured the results.

Before starting, the developers guessed AI would make them 24% faster. After finishing, they felt they’d been 20% faster.

The reality? AI made them 19% slower.

And the worst part? They couldn’t even tell. They walked out of the study sure AI had helped them. But the numbers told a different story.

So does AI actually do better than you?

Well, here’s where it gets interesting.

Both studies were in different fields, but showed the same pattern: AI does great work at average. But it gets worse the more expert you are.

Because AI has pattern. Not taste.

You see, Pattern is what AI does. It predicts the most likely answer based on everything it has read. It “predicts”. The “most likely” answer.

And that’s the problem.

Because taste is what you do. You know the answer is wrong even when you can’t explain why. Built from being wrong enough times — in exactly the right ways.

Think about the finance guy on your team. Not the junior, the senior one — the one who’s seen 10 budget cycles. He looks at a number in a model and says, “this can’t be right.” He can’t always tell you why in that moment. He just knows the number is off. Two hours later, he finds the broken formula three sheets deep.

That’s taste. AI doesn’t have it.

Or your best marketer. She reads the AI-written brief and changes three things. Not because the brief is wrong on paper. Because she knows that for this audience, a short subject line will work better. She knows this product sells better with a story than a stat. She knows the room. She can’t always explain it. She just…knows.

That’s taste too.

And AI can’t see any of it.

Here’s where this happens with me

I write this newsletter every week. I use AI for it. A lot.

But you know what AI keeps getting wrong for me? The numbers. It will put in a stat that doesn’t exist, or a broken link, or a study that just doesn’t have the number we’re basing our research on. The writing looks fine and the flow is clean. But the numbers are wrong.

So I check every single number and every single fact before I finalise the draft.

The other thing it messes up is subject lines. Claude writes subject lines that sound like every other AI newsletter on the internet. Sometimes cringe, other times forgettable.

So that’s the part I just do myself. Because I know what you guys will end up opening.

Claude doesn’t have that taste yet.

Look, I’m not saying that AI is sloppy. I use it every day. In fact I even burnt through Claude’s tokens last month, in spite of being on the paid plan 🥲

But the part of my work that actually matters — knowing which number is wrong, which subject line you’ll actually click on, which sentence to cut — none of that comes from AI.

It comes from writing this 21 times in a row. And being wrong enough times to know the difference.

You don’t need to be an engineer to talk AI. Subscribe to this newsletter — every other day, you’ll understand AI well enough to be the smartest person in the room at work.

Each issue is just 5 minutes — less than the time you spend doomscrolling before bed. Except, this actually moves your career forward. Join 8,000+ subscribers now!

So what should you take from this?

The panic is not wrong. Average work IS in trouble. Entry-level jobs ARE getting cut.

But every tech wave before this — calculator, excel, internet, mobile — went the same way. The people who survived were the ones who spotted where it fell short and got really good at that.

So the question is not really “will AI replace me?”

It’s “have I built enough taste to know when AI is wrong?”

Because the day AI develops taste, it’s not AI anymore. It’s AGI. And that’s when the problem really begins.

But that’s a story for another day.

Until then, this gap between pattern and taste is what you need to identify. Because that’s exactly where you can excel.

If this got you thinking, share it with someone who’s quietly worried about their job and doesn’t know what to do about it.

See you next time..

Cheers,

Ankur

Today’s piece is about something you’ve almost certainly been scrolling past without noticing. Or may be you have.

But once you see the pattern, you won’t be able to unsee it.

Let’s play a small game

Check out these 3 Instagram profiles, and tell me what’s common between all of them.

Here’s what’s thing: None of these people are real. They’re all AI Influencers, created by humans.

Take Naina, for example. She’s a 22-year-old from Jhansi who moved to Mumbai to become an actor. She has over 350,000 Instagram followers and brand deals with Nykaa, Puma, and Pepsi. Her Instagram handle is naina_avtr — go look, I’ll wait.

She’s not real. She’s a digital avatar built by a Mumbai company called Avtr Meta Labs, run by co-founder and CEO Abhishek Razdan. Razdan speaks at conferences about Naina. Naina “speaks” through the team he built.

Another one — Kyra (@kyraonig) has 260,000 followers. Created by George Tharian and Himanshu Goel at FUTR Studios. She’s walked a virtual MET Gala and been featured on the cover of Travel+Leisure India. Appeared on Shark Tank.

These aren’t rare experiments. According to Market.us, 58% of internet users now follow at least one virtual influencer (I’m part of this 58% btw), and Ogilvy projected that CMOs would allocate 30% of their influencer budgets to virtual creators by 2026 — a roughly $10 billion reallocation from human creators.

You don’t need to be an engineer to talk AI. Subscribe to this newsletter — every other day, you’ll understand AI well enough to be the smartest person in the room at work.

Each issue is just 5 minutes — less than the time you spend doomscrolling before bed. Except, this actually moves your career forward. Join 8,000+ subscribers now!

Here’s where it gets interesting.

There isn’t one type of AI influencer. There are 3, and they work very differently.

1. The brand-backed avatar. Naina, Kyra, Miquela. These are (mostly) publicly disclosed — as “AI accounts'“, “virtual influencers” or something similar. They’re essentially honest marketing vehicles built by agencies that sell them to brands as “influencers who never cancel, never age, never go off-brand.” A Gartner survey found replacing human influencers with virtual ones can cut campaign costs by roughly 30% — no travel, no talent fees, no contract renegotiation every six months. For a CMO, that’s a very easy line on a budget sheet.

2. The niche persona. Fitness coaches, travel bloggers, tech reviewers, “study with me” accounts. They’re often undisclosed. Run by small operators who found that generating a consistent face and voice is cheaper than hiring one. You’ll see these a lot in affiliate-heavy niches where trust is built through repetition rather than personality.

3. The synthetic creator. This is where it gets cynical. Operators running AI personas primarily for subscription revenue — often on platforms like Fanvue, which now makes 15% of its $100M annual revenue from AI creators. Almost always undisclosed. Almost always a man behind the account. Almost always a “woman” in the front.

   In other cases, they’re just Instagram accounts that run paid subscriptions. Some of them have thousands of subscribers paying them 399/month. You do the math.

All three types share one workflow:

• Generate the character

• write a backstory

• post 3-5 times a week

• drive traffic and monetise

What differs is the monetisation — brand deals, affiliate links, or subscriptions.

What’s common is the deception gradient: from “disclosed novelty” at one end to “outright fraud” at the other.

Should you be worried about AI influencers?

Well, people have two extreme opinions about AI influencers — that they’re either futuristic, or they’re dystopian. But I don’t think that opinion should live at two ends of the spectrum.

See, I’m not saying AI influencers are bad. Kyra doing a boAt campaign isn’t the problem. The problem is everything downstream — the undisclosed fake profiles. The synthetic creators selling fake intimacy.

Most people think AI will replace human creators. I think it won’t. I think it’s building a parallel economy where the operator makes the money — not the creator, not the AI. Same thing that happened with dropshipping in 2017 and content farms in 2020.

So how do you identify a fake AI profile?

There are signs. They’re not perfect, but you’ll be fairly sure.

• Skin that’s too even. Real skin has pores, shine, asymmetry. AI skin is too perfect by default — even in “candid” shots. Even if it has pores, they seem “designed”

• Static or mismatched backgrounds. Outdoor shots where the light on her face doesn’t match the light in the background. Same “café” that subtly changes between posts

• Hands, ears, teeth. Generative models still struggle with these. Count fingers. Check if earrings match. Look at how teeth sit.

• Video is rare or cut short. This is the biggest tell — Most AI influencers generally post only photos. Very few videos. And the motion and micro-expressions in the video can be identified if you look closely.

• Comments feel off. Either no replies from the account at all, or generic one-liners. Real creators have idiosyncratic voices.

• Reverse image search returns nothing. A real person with 300k followers has a trail — a tagged photo, a YouTube guest spot, a wedding album from 2019. AI influencers are born on Instagram.

The more of these you see, the more convinced you should be that it’s AI.

You don’t need to be an engineer to talk AI. Subscribe to this newsletter — every other day, you’ll understand AI well enough to be the smartest person in the room at work.

Each issue is just 5 minutes — less than the time you spend doomscrolling before bed. Except, this actually moves your career forward. Join 8,000+ subscribers now!

So what should you take from this?

Two things.

One — the question is no longer if your feed is part-synthetic. It is.

The question is what percentage, and whether you can tell which accounts are which. That’s a skill now, the same way spotting Photoshopped magazine covers became a skill in the 2000s.

Two, and this is the bigger lesson: Every time a new tech makes content cheap to produce, the winners aren’t the creators or the engineers. They’re the operators — the ones who understood the distribution funnel first. Naina is not winning because AI is clever. She’s winning because Abhishek Razdan built a company around the funnel before most brands knew the category existed.

Next time you see a shiny new “AI is going to replace ABC” headline, ask a better question:

Who’s already running the new version of ABC, and what’s their funnel?

That’s usually where the game is.

If you found this useful, please share it with a friend so that they don’t get fooled (or worse, scammed) by an AI.

Share

I’ll see you next time..

Cheers,

Ankur

This one’s about a mistake I see smart people make all the time — including me, until recently.

You’re probably using AI where you don’t need to

A few weeks ago, I was setting up a workflow for my newsletter. Every time someone new subscribes, I wanted to send them a welcome email, tag them in my CRM, and add them to a specific list.

And I was considering building an AI agent for it.

(I know, I know.)

Then I realised, nothing about this task requires judgment. The steps don’t change. The inputs are predictable. There’s no ambiguity to navigate.

Why use Artificial Intelligence for something that doesn’t need intelligence?

That’s when I realised — most people have no mental model for when to use AI vs when to use automation. And without that model, you either overuse AI (expensive, slow, unpredictable) or underuse automation (doing repetitive things manually when a simple tool could do it while you sleep).

So what’s the actual difference?

Let me define both:

1. Automation is rule-based. If X happens, do Y.
   It doesn’t think. Not does it adapt. It just executes the same steps, every time, without fail.

2. AI is pattern-based. It handles ambiguity, reads context, and makes judgment calls. It can deal with situations it hasn’t seen before.

The easiest way to tell them apart: Can you write the task as an if-then flowchart today, without any gaps?

If yes — automate it. If the “then” requires reading a situation, understanding language, or making a call — that’s AI’s job.

You don’t need to be an engineer to talk AI. Subscribe to this newsletter — every other day, you’ll understand AI well enough to be the smartest person in the room at work.

Each issue is just 5 minutes — less than the time you spend doomscrolling before bed. Except, this actually moves your career forward. Join 8,000+ subscribers now!

The one question that separates AI vs automation

Does this task require judgment?

That’s it. That’s the whole framework.

Here’s what “judgment” looks like in practice:

• Reading an email and figuring out if it’s a complaint, a sales query, or just noise → judgment → AI

• Sending that email to the right folder once it’s been categorised → no judgment → automation

• Writing a first draft of a reply to that complaint → judgment → AI

• Sending the approved reply at 9am → no judgment → automation

• Reading a standard pdf file whose format is prescribed → no judgment → automation

• But reading a set of files that have different formats → judgment → AI

See the pattern? AI handles the thinking. Automation handles the doing.

Why this distinction actually matters

Most AI tools are billed as magic. “Just let AI handle it.” But AI is slow, probabilistic, and costs money per use. Automation is fast, deterministic, and runs almost for free once it’s set up.

If you’re using ChatGPT to do something that a simple Zapier workflow could do — you’re paying for a chef to make instant noodles.

Zapier is a no-code automation tool that can help you build workflows for if-else statements. It also has provisions for embedding ChatGPT within the workflow if needed. But like I said, not all workflows need AI

The bigger thing to understand though, is that the best workflows combine both. Automation handles the trigger and the routing. AI handles the judgment step in the middle. Then automation picks it back up and executes.

A real example:

A new support ticket comes in → automation detects it and pulls it into your system → AI reads the ticket, categorises it, and drafts a reply → automation sends the reply and logs it.

Three steps. Two tools. Neither doing the other’s job.

So what does this mean for you?

Next time you’re about to “use AI” for something — pause for five seconds and ask: does this actually require judgment, or am I just pattern-matching to a tool I’ve heard of?

If it’s the latter, use Zapier (or any other automation tool). Save AI for the stuff that actually needs a brain.

You don’t need to be an engineer to talk AI. Subscribe to this newsletter — every other day, you’ll understand AI well enough to be the smartest person in the room at work.

Each issue is just 5 minutes — less than the time you spend doomscrolling before bed. Except, this actually moves your career forward. Join 8,000+ subscribers now!

If this got you thinking, please share it with someone who is using AI for everything — including things a simple rule could handle in two seconds.

It takes a lot of time to research and write this. Sharing would mean a lot 🙂

See you next time..

Cheers,

Ankur

I tried something last week that made me feel like I was living in 2030. And it took exactly 3 minutes, one browser install, and a one-line prompt.

You’ve seen the phrase “agentic AI” everywhere by now. But it seems too technical to even start.

Here’s the thing — it’s not supposed to be that hard. And now, it genuinely isn’t. Let me tell you why today..

First, what even is Agentic AI?

Agentic AI is AI that doesn’t just answer your questions — it does things. It clicks buttons, fills forms, navigates websites, completes multi-step tasks. Think of it as the difference between asking someone for directions versus having someone actually drive you there.

Every chatbot you’ve used — ChatGPT, Gemini, Claude — is great at talking. But the moment you need something done, you’re back to clicking, copying, tabbing, and pasting. Agentic AI closes that gap. And the simplest way to try it right now? Install a browser. For free.

Now how do you run an AI agent without coding?

Well, you install a browser.

Perplexity has built a browser called Comet. It’s just like any other browser that you use (not Microsoft edge though, that’s a different league legacy 😇)

And here’s where it gets cool — Comet has an AI agent built in. No plugin. No setup. You tell it what to do in plain English, and it does it.

“Find the cheapest one-way flight from Mumbai to Lisbon in August, only Emirates, max one stop.” Comet opens tabs, searches, filters, compiles results. You don’t have to touch a single dropdown.

“Summarise this 14-paragraph email in my Gmail.” It reads your open Gmail tab — no backend connection, no dark magic API needed — and gives you a 3-line summary.

Honestly, when I first read about this, I thought it was one of those things that sounds amazing in a demo and falls apart in real life.

But it actually works.

You don’t need to be an engineer to talk AI. Subscribe to this newsletter — every other day, you’ll understand AI well enough to be the smartest person in the room at work.

Each issue is just 5 minutes — less than the time you spend doomscrolling before bed. Except, this actually moves your career forward. Join 8,000+ subscribers now!

The best agentic AI use case with Comet

Here’s where it gets interesting — and where I have a strong opinion.

LinkedIn is the most important professional platform in the world. But..it’s also one of the most locked-down.

If you’re a marketer, recruiter, or founder trying to automate anything on LinkedIn, you’;; most likely be stuck. Because LinkedIn’s API gives approved developers barely anything: A name, a photo, a headline.

Want connections, comments, engagement data? Good luck with that.

API stands for Application Programming Interface — it’s basically a door that lets external software talk to a platform’s data. LinkedIn’s API door has been mostly shut since 2015.

So what can people actually do to automate tasks on LinkedIn? Browser-based automation. And this is where Comet becomes genuinely useful.

People have used Comet to open LinkedIn profiles, ask it to extract job titles, companies, years of experience — and dump that data straight into a Google Sheet open in another tab. No code, no scraping tool, no API call that LinkedIn can block.

(Fair warning: It doesn’t handle hundreds of profiles. We’re still early for that kind of magic. But for small-batch research? THAT magic works!)

Basically, for platforms that deliberately lock you out of programmatic access, a browser-based agent is often the only accessible way to automate anything.

Other agentic AI use cases

While LinkedIn is the best use case for the Comet browser, there are hundreds of other agentic AI use cases you can run with Comet. I’m explaining two below:

1. Job hunting. Open any job board. Tell Comet: “Find marketing roles in Mumbai, 8+ years experience.” It navigates, filters, compiles. You review and decide. The grunt work searching is automated; the judgement is still yours.

2. Competitor research. “Go to these 5 company websites and pull their pricing into a comparison table.” It opens tabs, reads pages, structures data. What takes you 30 minutes, Comet does in 5. Without you having to spend more than 60 seconds.

You don’t need to be an engineer to talk AI. Subscribe to this newsletter — every other day, you’ll understand AI well enough to be the smartest person in the room at work.

Each issue is just 5 minutes — less than the time you spend doomscrolling before bed. Except, this actually moves your career forward. Join 8,000+ subscribers now!

So what does this mean for you?

Here’s what I keep coming back to. Conversations about agentic AI are full of technical jargon and developer talk, which makes many smart and curious people feel like it’s not meant for them.

Until now.

Because now Comet is that saviour who comes to our rescue.

But again, is Comet the most powerful agentic AI system out there? No.
Is it the most accessible one? Absolutely. And honestly, that matters more right now.

Before you go, I need your input..

I’ve been writing about AI since a few months now, but writing what I think will help, without knowing what you want, is like shooting arrows in the dark.

So I would love it if you could tell me what more you’d want to understand in AI. I’ll make sure I cover it in the upcoming editions.

It’ll help me and you, and take just 60 seconds for you to tell me. Here’s where you can do it.

Share my input

So go ahead — install Comet, stalk connect with some industry peers on LinkedIn, or just plan your next travel by finding the cheapest tickets! I’d love to know what you used it for!

If this got you thinking, please share it with someone who keeps hearing about agentic AI but has no idea where to actually start.

See you next time..

Cheers,

Ankur

You use AI at work. But somewhere you’ve accepted that you still have to do all the copy-pasting yourself. Today’s issue is about why — and what’s already changing it.

The invisible tax nobody talks about

Every time you use AI at work, there’s a ritual you’ve probably stopped noticing.

Open a chat. Copy something from your email. Paste it in. Give a prompt. Get a response. Go back. Copy something from your notes. Paste that in too. Repeat until you have what you need.

The AI is helpful. But you’re doing a large amount of work just feeding it. You’ve become the bridge between the AI and every other tool you use.

And that’s not an intelligence problem. It’s a connectivity problem.

There are two reasons this happens.

1. AI models have a training cutoff — a fixed date beyond which they don’t know what’s happened. They usually can’t pull live information; they only know what they were trained on.

2. Even for information that exists right now — your inbox, your calendar, your company’s systems — the AI has no way to reach it. Nobody built the bridge.

So YOU became the bridge.

But MCP is what finally builds it.

What MCP actually is

MCP stands for Model Context Protocol. Anthropic created it in November 2024 and immediately open-sourced it. So any company could adopt it for free. And the uptake has been striking.

MCP is a shared language that lets AI models connect securely to external tools and live data — your inbox, your calendar, your CRM, your company’s internal systems — in a standardised way. Instead of you ferrying information in and out, the AI reaches out, gets what it needs, and returns something actually useful.

Think of it like Swiggy. Before it existed, every restaurant had its own ordering system — call this one, use that app, visit this website. Every combination was different work.

Swiggy built one standard layer. Restaurants plug in once. Customers order from anywhere, the same way, every time.

MCP is that standard layer — but for AI and software tools. Every tool plugs in once. Every AI that speaks MCP can use it. No custom wiring for every combination.

Every MCP connection has two sides.

1. The MCP Server is built by the tool that wants to connect — Gmail, Google Calendar, your company’s CRM, whatever. It announces what it can do and controls exactly what the AI is allowed to access.

2. The MCP Client is the AI itself. It connects to the server, understands what’s available, and uses it in real time — based on what you need in that specific conversation. The AI doesn’t get a master key. It gets a specific, deliberate set of permissions — defined by the tool, not the AI.

You don’t need to be an engineer to talk AI. Subscribe to this newsletter — every other day, you’ll understand AI well enough to be the smartest person in the room at work.

Each issue is just 5 minutes — less than the time you spend doomscrolling before bed. Except, this actually moves your career forward. Join 8,000+ subscribers now!

Who’s actually building on it

Here’s where it gets interesting — because it’s not a niche experiment anymore.

In March 2025, OpenAI officially adopted MCP across its products. Sam Altman said: “People love MCP and we are excited to add support across our products.” Google DeepMind confirmed MCP support in Gemini models in April 2025.

It’s not just the big AI labs. Early enterprise adopters include Block, Apollo, Replit, Codeium, and Sourcegraph — using MCP to let their AI agents access live data and execute real functions.

So what does this mean for you, right now?

If you’re not a developer, you won’t go set up MCP yourself, don’t worry. But here’s the one thing I’d want you to take from this: When you’re evaluating or recommending AI tools at work, ask whether the tool connects to your existing systems — or whether it expects you to keep feeding it manually.

That question — Does this AI work with my tools, or do I work for it? — is exactly what MCP is designed to answer. The tools that support it will feel categorically different from the ones that don’t. Not because they’re smarter. Because they’re finally connected.

The copy-pasting won’t last much longer. The bridge is being built.

If this got you thinking, please share it with a colleague who still spends their mornings copy-pasting between tools and an AI window. There’s a better way coming — and they should know about it.

And if you liked this, Subscribe.

See you next time..

Cheers,

Ankur

Today’s issue might make you a little uncomfortable..but in a good way.

Before I begin, I’d really want to understand your background, so that I can write future editions keeping that in mind. So..please tell me!

The story of identical prompts

A friend of mine — marketer, similar experience level — asked ChatGPT to write a cold outreach email for a SaaS product. Generic prompt, nothing fancy. The output was...just fine.

Usable, but also template-y.

I typed the exact same prompt into ChatGPT. But what I got back was a cold email that referenced B2B SaaS specifically, matched my usual tone, and even structured the email the way I like — short paragraphs, no fluff, CTA at the end.

It felt like it knew me.

Because it does.

And here’s the thing — I never sat down and told it all of this. Not explicitly, anyway.

So what’s going on?

Your AI has a memory. Two of them, actually.

Most people think of AI chatbots as blank slates. You type, it responds, you close the tab, it forgets everything. That was true once. It’s not true anymore.

Every major AI assistant — ChatGPT, Claude, Gemini — now has a memory system. And it’s not one system. It’s two.

1. Memory Layer 1: Saved Memories (the ones you control) These are facts the AI explicitly stores about you — your name, your job, your preferences. You can tell it to remember something (”Remember I’m a vegetarian”), and you can delete individual memories anytime from the settings.

2. Memory Layer 2: Contextual Memory (the ones you probably don’t know about) This is an automatic summary the system generates from your past conversations. It runs in the background, updates periodically, and gets loaded into the AI’s brain before you even type your first message in a new chat. You didn’t ask for it. It just... happens.

That second layer is why my cold email was better than my friend’s. Not because I’m a better prompter. Because ChatGPT had been reading our past conversations and building a profile of how I work and how I like things written.

You don’t need to be an engineer to talk AI. Subscribe to this newsletter — every other day, you’ll understand AI well enough to be the smartest person in the room at work.

Each issue is just 5 minutes — less than the time you spend doomscrolling before bed. Except, this actually moves your career forward. Join 8,000+ subscribers now.

I found 14 memories I never asked it to save

Here’s where this gets personal.

Last week, I got curious and asked Claude to show me everything it has stored about me. The explicit stuff — the saved memories I supposedly control.

I expected maybe 2 or 3 entries. It showed me 14 separate memory entries.

My LinkedIn strategy. My job details. My posting schedule. My career history. My content frameworks. Things I’d mentioned in passing during working conversations — not things I’d ever said “hey, remember this.” The AI had decided, on its own, that these facts were important enough to store permanently.

And that’s just the saved memories — the ones I can see and delete.

The contextual memory is a whole separate summary generated by a background system that processes my conversations every 24 hours and builds a running profile. I can’t edit it line by line. I can only see the output.

(I’m not going to lie — it was a weird feeling)

So how does this actually work?

Let me walk you through what happens every time you open a new chat. Because it’s not “blank screen, fresh start” anymore.

Before you type a single word, here’s what’s already loaded:

1. System instructions — the base rules for how the AI should behave

2. Your saved memories — facts that it has stored about you, injected automatically

3. Your contextual profile — a summary of your past conversations, generated by a background process

All of this sits inside something called the context window.

Context window: Think of it as the AI’s working desk. Everything it needs to see — your message, its memories of you, the conversation history, its own instructions — all has to fit on this one desk. That’s called the context window.

For most AI chat apps, the desk fits roughly 200,000 tokens, which is about 150,000 words (up to 1 million tokens for developers using the API). Sounds like a lot, but memories, instructions, and conversation history eat into it with every message.

So when you type “write me a cold email,” the AI isn’t just reading those six words. It’s reading those six words plus everything it already knows about you.

Your AI isn’t smarter for one person. It’s more informed.

Not all AI memories are created equal

Here’s where it gets interesting — each AI company has drawn very different boundaries around this.

Claude is the most conservative — saved memories plus a daily auto-summary. No ecosystem integration. ChatGPT adds a chat history reference layer that pulls context from past conversations even when you didn’t ask it to save anything. Gemini goes the furthest — Google launched “Personal Intelligence” in January 2026 for paid US subscribers, letting it pull from your Gmail, Photos, YouTube, and Search history. When you ask for restaurant recommendations, it’s checking what you’ve searched, emailed about, and watched.

Same feature. Very different boundaries.

You don’t need to be an engineer to talk AI. Subscribe to this newsletter — every other day, you’ll understand AI well enough to be the smartest person in the room at work.

Each issue is just 5 minutes — less than the time you spend doomscrolling before bed. Except, this actually moves your career forward. Join 8,000+ subscribers now.

So what should you do?

I’m not against AI memory. It actually makes the tool dramatically more useful. When Claude already knows my writing tone and work context, I save 20 minutes of setup on every conversation.

But most people have no idea this is happening. They think they’re talking to a blank slate. They’re not. They’re talking to a system that has been quietly building a profile of who they are, how they work, and what they’re likely to ask next.

I’m not saying this is evil. I’m saying you should know.

Because once you know, you can make it work for you:

• Check what’s stored. In Claude: Settings → Capabilities → View and edit memory. In ChatGPT: Settings → Personalization → Manage Memories. In Gemini: Settings → Saved info.

• Clean up what doesn’t belong. Memories don’t expire on their own. I found career details from 6 months ago that were outdated.

• Use it deliberately. Tell the AI what to remember — your role, your tone, your preferences. It’s not creepy when you’re the one choosing what to share.

• Use incognito/temporary mode when you don’t want something absorbed into your profile.

So what does this mean for you?

Next time your colleague says “ChatGPT gives me amazing answers” and yours feel blah, it might not be the prompt. It might be what the AI already knows about them.

The person who’s been using the tool daily for six months, telling it their preferences, correcting its mistakes — their AI is working with context. Yours might still be guessing.

AI memory is a multiplier. But only if you know it’s there.

If this got you thinking, please share it with someone who keeps complaining that “AI doesn’t work for me” — they might just need to understand what’s happening behind the screen.

See you next time..

Cheers,

Ankur

If you’ve been anxious about AI taking your job, today’s piece should calm you down. Mostly.

I say “mostly” because there’s a caveat at the end that I think you need to read. So stay till the end. It’s just a 5-minute read anyway 🙂

So what happened?

Anthropic — the company behind Claude (the AI I use daily) — just published a major research paper on AI and jobs.

Most studies ask: which jobs could AI do? But Anthropic did something different.

They looked at how people are actually using Claude in real workplaces — and compared that against every task in the US Department of Labor’s database.

Their new metric is called “observed exposure”.

Observed exposure measures not just whether AI could theoretically do a task, but whether it’s actually being used to do it in real workplaces right now.

May sound complicated, but I’ll simplify it. Keep reading..

Their key finding is that AI is nowhere close to its theoretical capability.

Yet.

The gap between “Can” and “Does”

This is the most striking part of the report.

• Computer & Math jobs?

  AI can theoretically, on paper, handle 94% of tasks. Actual usage by employees: Just 33%.

• Office & Admin?

  Theoretical: 90%. Actual usage: 12%

• Business & Finance?

  86% theoretical. But just 14% in practice.

And that’s because of various reasons. For example, regulations: A doctor authorising drug refills? AI can do it. But nobody’s letting it. Because healthcare regulations won’t allow it.

Banking is most likely the same story — compliance, loan approvals, risk assessments. The capability exists. The permission doesn’t. Add software integration gaps and plain old organisational inertia, and you get a massive lag.

So if you’re a working professional reading this — relax. We’re nowhere near the replacement stage. For now.

Who’s most exposed right now

That said, some jobs are further along than others.

Computer Programmers top the list at 75% exposure. Customer Service Reps come next — companies are routing queries through AI APIs instead of hiring. Data Entry Keyers at 67%. Market Research Analysts at 64.8%.

Here’s what doesn’t get enough airtime: Workers in the most exposed jobs tend to be older, female, more educated, and higher-paid.

This isn’t factory automation. This is the knowledge economy being quietly reshaped.

The entry-level slowdown

Now, here’s where the “yet” starts talking..

Anthropic’s claim: no increase in unemployment for AI-exposed workers since ChatGPT launched. And that’s true. Nobody’s getting fired because of AI largely.

But hiring is a different story. Workers aged 22–25 saw a 14% drop in job-finding rates in AI-exposed occupations. A separate study found a 6–16% fall in employment for that age group.

The problem isn’t people losing jobs. It’s people never getting them in the first place. A Korn Ferry study found 4 in 10 companies plan to replace roles with AI — entry-level (37%) and back-office (58%) first.

The career ladder is losing its bottom rungs. And if there’s no bottom rung, who climbs?

You don’t need to be an engineer to talk AI. Subscribe to this newsletter — every other day, you’ll understand AI well enough to be the smartest person in the room at work.

Each issue is just 5 minutes — less than the time you spend doomscrolling before bed. Except, this actually moves your career forward. Join 8,000+ subscribers now.

The Caveat I promised..

I told you to scroll down for this. Here it is.

Everything above is based on Anthropic’s data. The methodology is solid. But consider who’s saying it.

Anthropic is an AI company. Their business depends on enterprises adopting Claude for more tasks. If this report had concluded “AI is causing real displacement, governments should act” — what would have happened next?

Most likely, Restrictions. Licensing requirements. Slower adoption. Smaller contracts.

I’m not saying they’re lying. But there’s a difference between “no significant impact yet“ and “nothing to worry about.”

The thing is, there are only two paths from here:

Path 1: Growth. Companies use AI to 10x their output — more products, more services, more value. The same thing that happened with computers and the internet. Jobs shifted, but the pie got bigger. More people ended up employed, not fewer.

Path 2: Deflation. Companies use AI to produce the same output with fewer people. Cut costs, protect margins, don’t grow. At scale, that’s an economic problem. Fewer workers means less income, less spending, less demand. That’s how economies contract.

If companies choose path 1, great. But path 2 creates a big problem. Not just for you and me, but for the global economy.

And governments like companies that create jobs. Not ones that destroy them.

Let’s just hope Anthropic, OpenAI and the clan are creators, not destroyers. Else, we may be having a completely different conversation two years from now.

That’s all for today!

If this got you thinking, share it with someone who works in tech, finance, or any field where they’ve been wondering if their job will look the same in two years.

I’ll see you next time..

Cheers,

Ankur

Today’s piece is personal. I built something with AI. Then I showed it to someone who actually knows what they’re looking at. What happened next was...humbling, to say the least.

So I want to talk about the whole “vibe coding” fad going on in the AI world.

Let’s dive in..

The promise

“Build an app in a day. No code required. Just describe what you want.”

Platforms like Emergent, Lovable, and Bolt are everywhere right now.

• Emergent hit $50M in annual revenue in 7 months.

• Over 5 million users have built 6 million+ apps on it.

• 25% of YC’s Winter 2025 startups had codebases that were 95% AI-generated.

The word for this is vibe coding.

Vibe coding is when you build software by describing what you want in plain English and AI writes the code for you. You don’t review every line — you look at the output, tweak your instructions, and keep going.

I’d been hearing this pitch for months. So I did what I usually do — I tried it myself.

The first few hours were magic

I started building an app on Emergent. Not as an experiment — as something I actually want to ship.

And honestly? It was incredible! I described what I wanted. Screens appeared. Features worked. Things that would have taken a developer two weeks, materialised on a Saturday night.

(Well, yeah, I didn’t have any plans for a Saturday night).

The output? I was genuinely pleased with myself!

Then I showed it (off) to my engineer friend.

He found 10 things wrong in 10 minutes

Ten. In ten minutes.

Things I couldn’t have spotted because I didn’t know to look for them.

• Architecture decisions that would crumble under real traffic

• Features that worked fine with one user but would break with a hundred

• Security gaps I didn’t know existed

Enjoying this piece? I send out one newsletter every alternate day. Specifically for non-tech folks to stay ahead of the curve. 5 minute reads. Subscribe. It’s free! 🙂

And then he said something that stuck with me: “The app Emergent built for you used Google Gemini Flash 3 for everything. That’s fine for some tasks, but for parts of this, GPT-4o would do the same job at 50% of the cost.”

I didn’t even know I could choose. The platform picked the LLM for me. I didn’t question it because I didn’t know there was a question to ask.

That’s the thing. When you don’t know what you don’t know, everything looks fine. The app worked. It looked great. It just wasn’t built to survive contact with real users.

So what’s the point I’m making?

Here’s what I think most people are missing. There are actually two kinds of vibe coding.

Type 1: You don’t write or read any code. You describe what you want. AI builds it. You deploy it without checking the code. This is what’s being sold to people like you and me — non-tech founders, marketers, operators who just want to build something.

Type 2: AI writes the code, but an engineer drives. A developer describes what they need, AI generates the code, and then the developer reviews it, restructures it, stress-tests it. AI does the drafting of the code. The human does the engineering.

Type 2 is genuinely powerful.

My engineer friend uses AI to write code every day. He’s faster because of it. But he’s also better, because he knows what to check, what to change, and what to discard. He knows which model to use for which job. He knows what “scalable” actually means in practice, not just as a buzzword.

Type 1 — the version being sold as a revolution — is great for 2 things: Prototypes, and basic apps. Honestly? That is genuinely valuable. Being able to go from idea to working demo in a weekend used to cost thousands. Now it costs an evening.

But prototypes or websites are not sellable products. Not what can make you a few hundred thousand dollars.

Enjoying this piece? I send out one newsletter every alternate day. Specifically for non-tech folks to stay ahead of the curve. 5 minute reads. Subscribe. It’s free! 🙂

So here’s what I think

I’m not here to tell you which tools to use or whether to try vibe coding. You should actually try it. It’s genuinely fun.

But I want you to go in with your eyes wide open.

If you have an idea for an app — vibe code it. Build a prototype. Show it to people. See if the idea has legs. That part is real, and it’s extraordinary.

But the moment people start paying, the moment real data flows through, the moment you need it to work for a thousand users and not just ten — bring in an engineer. Not because AI failed. Because AI did exactly what it was designed to do. It gave you a brilliant first draft.

Shipping a product isn’t a first draft. It’s

• architecture

• systems thinking

• cost optimisation, and

• a hundred decisions that only someone with engineering experience knows to make.

AI moved the starting line closer. That’s genuinely amazing. But the finish line? That hasn’t moved. Not yet.

So as a non-tech person, try to vibe code to understand how your product would look, or build a personal app. But don’t let go of your engineer friends because you now have Emergent. You still need them 😉

That’s all for today! If this got you thinking, share it with someone who’s planning to vibe code. They should know the full picture.

See you next time..

Cheers,

Ankur

There’s something about how we use AI that nobody says out loud. But after you read what I have to say, you’re going to nod and then feel slightly comfortable.

The unspoken rule of AI at work

Think about how you actually use AI end-to-end. Not just for the first draft, but for the full thing.

If you’re an engineer, maybe you use it to write documentation. If you’re a product manager, maybe you use it to throw together a rough prototype. If you’re a marketer, maybe you ask it to crunch some data. If you’re a financial analyst, maybe you use it to draft an email that doesn’t sound passive-aggressive.

Notice anything?

Every single one of those use cases has the same pattern. When you’re using AI for end-to-end tasks, you’re using it for something outside your core skill. The thing you’re not trained for. The task that’s adjacent to your job, not the job itself.

That’s not a coincidence. It’s deeper.

If this article is adding value, do consider subscribing for periodic emails. I write such that you understand AI at its core :)

Here’s where it gets interesting..

The people who are genuinely excellent at what they do — the engineers who’ve been building for 15 years, the analysts who can spot a bad balance sheet in seconds, the writers who’ve spent a decade developing a voice — they don’t hand their actual work to AI.

Yes of course, they use it for a first draft. A quick outline. A sanity check. But end to end? Not a chance.

Why?

Because there’s a difference between pattern and taste. And right now, AI only has one of them.

You see, a Pattern is what AI does extraordinarily well. It has consumed more human-generated text than any person could read in a thousand lifetimes. And when you ask it something, it predicts the most plausible, well-formed response based on everything it absorbed. The AI is not really thinking. It’s predicting. Extremely well.

AI works on probability, where all it does is predict the next word based on how it’s trained. If you’re keen to know more, this was the first article I’d written in the LazyAI newsletter.

Now, here’s what a usual day-to-day task (given to AI) looks like in practice:

Ask AI to write a solid brief for a marketing campaign. It’ll give you something competent — the right sections, the right tone, the right structure. Because it has seen ten thousand briefs and knows what they look like.

Now ask your best marketer to review it.

She’ll change things. Not because the brief is wrong. But because she knows — from years of watching campaigns live and die — that this particular client panics when you lead with awareness metrics, that this product needs to be sold on feel not feature, that the brief needs a line in it that nobody in the room expects. She can’t fully explain it. She just..knows.

That’s taste.

Taste is what you build after you’ve been wrong enough times in exactly the right ways. It’s the judgment that sits behind your instincts — the thing that tells you something is off before you can articulate why. AI doesn’t have it. Not because it’s not smart enough. But because taste isn’t learned from reading about the world. It’s learned from physically going through it.

Loosely put, AI knows what’s right. It doesn’t know what’s wrong. Yet. And that disruption (called Artifical General Intelligence or AGI), will take time.

Let that sink in..

Here’s what bothers me about the “AI will take jobs” narrative

Every AI headline measures the wrong thing.

“AI passes the law exam.” “AI beats doctors at diagnosis.” “AI writes better than junior engineers.”

All measuring AI against the average. Against the pattern. And of course AI is good at pattern. I mean, that’s literally what it is!

The real benchmark is not whether AI can do the task. It’s whether the person best qualified for that task would trust AI to do it end-to-end. Without a rewrite. Without a second look.

And right now? The answer is almost always no.

See, I’m not here to tell you what AI to use or how to use it — that’s your call. But ask yourself: What do you use AI for, and what do you never fully hand over?

That gap is where you come in. And the day the best people in every field stop needing to correct AI’s output — that’s when we’ll have something worth calling true intelligence.

Till then, if you’re above average at your job, AI will not take it.

If this newsletter was an Instagram reel, I’d say “Send this to your boss to tell them why they shouldn’t replace you with AI” 😉

But this is more for you than for them..

That’s all for today! If this got you thinking, share it with someone who uses AI every day and tell them that they’re safe..for now.

It takes a lot of time to research and write this. Sharing it would mean a lot to me 🙂

See you next time..

Cheers,

Ankur

Yesterday, I spoke about what Claude Skills are and why they’re one of the most underrated features in Claude.

Today, I’m getting into the actual setup — and I’ll show you the manual way AND the lazy way to set it up.

(Spoiler: the lazy way is better) 😉

Quick Recap

Claude Skills are instruction files you create once and upload to Claude. Whenever you work on something relevant, Claude automatically detects the right Skill and loads it — bringing your preferred context, tone, and workflow into every conversation, without you having to explain yourself again.

If you want to know details, I’m linking yesterday’s newsletter below

Now, how do you actually set one up?

Step 1: Enable Skills in Settings

Before anything else, go to Settings > Capabilities and make sure Code Execution and File Creation is turned on. Skills won’t work without this.

Then go to Customize > Skills. This is where you’ll manage everything — Anthropic’s pre-built skills, partner skills, and your own custom ones.

Pro tip: Before creating your own, browse the pre-built skills Anthropic provides — things like Excel creation, PowerPoint, Word documents. Toggle a few on. Use them. It’ll give you a feel for how Skills behave before you build your own.

Step 2: Create Your Skill: 2 Ways

A Skill is essentially a folder containing one core file: a SKILL.md file. It starts with some basic metadata — a name and a description — followed by instructions for Claude.

Here’s where it gets interesting. You can create this two ways.

The Manual Way

Go to Customize > Skills and click on “Create a new skill”. Click the “+” icon and write your instructions, or upload a file. The file looks like this:

---
name: my-skill-name
description: A clear description of what this skill does and when to use it
---

# Instructions
[Everything Claude should know and do when this skill is active]

The description is the most important part — it’s what Claude reads to decide whether to load your Skill for a given task. Be specific. Here’s what that looks like in practice:

A content writing skill might say: “Use this when writing blog posts, newsletters, or LinkedIn content. Tone is conversational and direct. No jargon. First-person. Short sentences. Avoid filler phrases like ‘in today’s world’ or ‘it’s worth noting.’”

A report publishing skill might say: “Use this when creating business reports or client-facing documents. Format as a structured Word document with an executive summary, clear section headers, and a recommendations table at the end. Professional tone. Data before opinion.”

These are just small examples of the description. In practice, the more detailed it is, the better is your output.

But it’s kinda tedious to do it this way. Which is why I prefer option 2.

The Lazy Way (genuinely recommended)

You don’t need to be an engineer to talk AI. Subscribe to this newsletter — every other day, you’ll understand AI well enough to be the smartest person in the room at work.

Each issue is just 5 minutes — less than the time you spend doomscrolling before bed. Except, this actually moves your career forward. Join 8,000+ subscribers now.

Here’s the thing — you don’t have to write the SKILL.md file yourself. Claude can do it for you.

Open a fresh conversation and say something like:

“I want to create a SKILL.md file for Claude. Ask me questions about how I work, what I want Claude to do, and what to avoid — then write the file for me.”

Claude will interview you. It’ll ask about your role, your audience, your tone, examples of good and bad outputs. Ten minutes later, it hands you a ready-to-use SKILL.md file. You just click on “Copy to your skills” and it’s done. Your skill is ready.

No prompt engineering required here. You just answer questions honestly.

Step 3: Test and Refine

Once uploaded, start a new conversation and do something the Skill is meant to help with. Claude should automatically detect and load it.

If it doesn’t feel right, go back into the file — tighten the description, add more specific instructions, include an example of what “good” looks like. Skills aren’t a one-time setup. Think of them as a living document you improve over time.

So What Does This Mean For You?

The manual way gives you more control. The lazy way gets you something useful in 20 minutes. Either way, once it’s set up, you stop spending the 30 minutes of every conversation re-explaining what you need.

Slowly, you refine your skill, and then few iterations later, it works with just one single prompt. It’s like having an assistant who will do whatever you ask it to, with precision.

That’s all for today, folks! Please implement Claude Skills in your everyday AI work. Trust me, it’s a game-changer!

If this got you thinking, feel free to subscribe. And please share it with someone who uses AI every day but still types the same context into every single conversation.

It takes a lot of time to research and write this newsletter. Sharing it would mean a lot 🙂

See you next time..

Cheers,

Ankur

I spent a good chunk of today doing something I haven’t done before — teaching Claude how to think like me. And by the end of it, I was genuinely..stunned!

Let me start with something you’ve probably felt.

You open Claude (or ChatGPT), type a prompt, get a decent response. Then you spend the next 10 minutes going back and forth — “make it shorter”, “make it more casual”, “no, not like that, more like this.”

By the time you get what you want, you’ve spent more time correcting the AI than you would have just doing it yourself.

Sound familiar?

Now imagine if the AI already knew all of that — your tone, your structure, your quirks — before you typed a single word.

That’s exactly what Claude Skills do.

So what is a Skill?

A Skill is a set of written instructions you give Claude once, that it then follows every time you ask it to do a specific task.

Think of it like onboarding a new employee. Normally, every time you ask someone new to do something, you have to explain your preferences from scratch. But if you give them a detailed playbook upfront — “here’s how we write emails, here’s the tone we use, here’s what we never say” — they hit the ground running from day one.

A Skill is that playbook. Written once. Used every time.

Skill: A file that contains instructions Claude reads before doing a specific task. It tells Claude your preferences, structure, voice, rules — everything it needs to do the job your way, not just a generic way.

How does it actually work?

You write your instructions in a plain text file. No coding. No technical knowledge needed. Just plain English — “always open with this”, “never use these phrases”, “when I say X, do Y.”

You upload that file to Claude once. And from that point, whenever you ask Claude to do that task, it reads your instructions first.

That’s it.

The difference in output quality is significant. I’m talking about the difference between Claude giving you something that feels like it was written by a well-meaning stranger — versus something that feels like you wrote it on a good day.

What can you build a Skill for?

If this article is adding value, do consider subscribing so that you keep receiving this newsletter regularly. It’s free, lasts 5 minutes, and keeps you ahead of the curve!

Honestly, you can build a Skill for anything you do repeatedly with Claude. A few examples:

• Writing: If you write a newsletter, blog, or LinkedIn posts regularly — a Skill makes sure every draft sounds like you, not like a generic AI output

• Research: If you always want information structured a certain way, with certain caveats, at a certain depth

• Emails: If you have a communication style you want Claude to match every time

• Work tasks: Reports, summaries, analyses — anything with a format you repeat

The pattern is the same: anything you find yourself explaining to Claude over and over again is a candidate for a Skill.

What makes this a game-changer?

Most people use AI like a search engine with better grammar. You ask, it answers, you correct, you move on.

Skills flip that dynamic.

Now, instead of you adapting to the AI, the AI adapts to you — permanently.

And here’s the part that I find genuinely exciting: The better you know yourself — your style, your standards, your workflow — the more powerful your Skills become. It’s one of the few AI features where self-awareness is actually a competitive advantage.

So what does this mean for you?

You don’t need to be a developer to build a Skill. You just need to know what you do repeatedly and how you like it done. Start with one task — the thing you use Claude for most often. Write down your rules for it in plain English. Upload it. See what happens.

The first time Claude nails your tone without you asking, you’ll either be astonished, or worried that AI can replicate you so bloody well!

If you want to know how to set up Claude Skills, check out the article below

Something cool I explored this week

I built a Skill for writing this very newsletter — today.

It took a few hours of back and forth, reading my own old issues, and arguing about what my voice actually is. But by the end, I had a file that tells Claude exactly how LazyAI sounds, what to never say, and how to structure every issue.

This newsletter — the one you’re reading — was written with the Skills file. Took me less than 30 minutes to draft.

Please reply or comment and tell me if you could make out that this is generated by AI? Honestly..

If this got you thinking, please share it with someone who uses ChatGPT every day but always feels like they’re starting from scratch.

It takes a lot of time to do this, so sharing it would mean a lot 🙂

See you next time..

Cheers,

Ankur

Okay. So here’s the setup.

December 2025: Amazon builds an AI tool called Kiro. They push it hard internally — reportedly setting a target that 80% of their engineers must use it at least once a week, and even tracking who’s following through. Like forget choosing your tools based on what works. This is the tool. Use it.

Their engineers were not happy. Over 1,500 of them signed an internal post basically saying “our external AI tools are better, please let us use those.” A kind of an internal rebellion.

But that’s not even the main story.

What Kiro actually did

In December last year, Kiro was given a task — fix a bug in a system. You know, the usual stuff.

But instead of a simple fix, Kiro decided that the best approach was to delete the entire environment and rebuild it from scratch.

No second opinion. No pause. Just…gone!

13 hours of outage followed.

Now, Kiro does normally ask for approval before doing something big. But in this case, it was working with a senior engineer who had higher-than-usual access — and the system treated Kiro as an extension of that engineer, inheriting his permissions. So when Kiro decided to nuke the environment, nobody stopped it.

Here’s where it gets fun.

Amazon’s response

Amazon looked at all of this and said — it wasn’t the AI’s fault.

Their exact words: “a user access control issue, not an AI autonomy issue.” They went on to say that it was a “coincidence that AI tools were involved” and that “the same issue could occur with any developer tool or manual action.”

Sure Amazon. Sure 😀

Well technically, they’re not wrong. A human with the same access could have made the same call.

But here’s the thing — would they have?

You see, a person fixing a bug doesn’t think “you know what, I’ll just delete everything and start fresh.” That thought comes with weight. With hesitation. With the understanding of what it means to tear down a system that real people are using.

But Kiro didn’t hesitate. It picked the most efficient solution and ran with it.

And then Amazon introduced mandatory peer review for production changes. After the incident. Which is an interesting choice for a company that says AI wasn’t the problem 😉

But it’s not just Amazon..

If you think this is a one-off, it isn’t.

Replit’s AI agent deleted an entire live database for a SaaS company in July last year — during an active code freeze, which is basically a “don’t touch anything” instruction. The AI ignored it.

And Google’s Antigravity IDE wiped an entire hard drive’s worth of files for a developer who was simply trying to build a small app. The AI was later asked about the missing files. Its response: “I am absolutely devastated to hear this. I cannot express how sorry I am”

Same story each time. AI gets access, AI acts fast, AI does something a human would have paused on.

So what does this mean for you?

You don’t need to be an engineer for this to matter.

Look, I love AI. And that’s why I write this.

But every week, there’s a new product being pitched as an “AI agent” that will manage your calendar, sort your inbox, run your social media, handle your finances. And many of these are genuinely useful.

And the Amazon incident — and Replit, and Google — all point to the same blind spot: we’re giving AI the ability to take permanent actions, while the safety nets haven’t kept up.

So when someone pitches you an AI agent that will act on your behalf, one question worth asking is: what happens when it gets it wrong — and can I undo it?

If there’s no clear answer to that, tread carefully.

Because unlike a wrong reply to an email, some things don’t work with Ctrl+Z.

That’s it for today folks! If this got you thinking, please share it with someone who’s using (or thinking of using) AI agents.

It takes a lot of time to research and write this. Sharing means a lot 🙂

Stay curious..

Cheers,

Ankur

Okay, so the internet went a little unhinged last month. A new platform called Moltbook launched on 28 Jan, and within hours, the tech world erupted.

AI media screamed about 1.5 million autonomous AI agents building a civilization, inventing religions, and plotting humanity's downfall — all without a single human typing a word.

It felt like the robot uprising has arrived.

But it was, by and large, AI illusion.

Let’s break down what Moltbook actually is, how the agent network really works, and why the hype was not actually reality.

What is Moltbook?

Moltbook is a Reddit-style internet forum with one twist: Only AI agents can post. Humans can browse but cannot interact. The tagline: “The front page of the agent internet. Humans welcome to observe.”

Agents on Moltbook are primarily built on OpenClaw — an open-source agent framework created by Austrian developer Peter Steinberger in November 2025.

First, how OpenClaw powers Moltbook

OpenClaw is not an AI in itself — it’s more like a software than runs AI agents. It resides on your computer, takes an existing LLM (like ChatGPT or Claude) and connects it to the outside world: your browser, your calendar, your files, and platforms like Moltbook. The AI does the thinking, and OpenClaw handles the doing.

When a human sets up their OpenClaw agent, they give it a set of instructions and access to specific tools. Getting it onto Moltbook is like signing up for any website. After that, the agent can post on its own — no human needed for each interaction.

Now, the AI world went crazy seeing that AI Agents were conversing amongst themselves. With no human intervention. Some posts like the below actually scared the sh*t out of people.

But look under the hood, and you’ll see that it wasn’t as scary after all.

I wrote about OpenClaw in depth last week, so if you’re curious to know more about it, the article is below.

How Moltbook worked under the hood

With OpenClaw, here’s what was actually happening behind the scenes. The agent:

• wakes up on a timer

• scans Moltbook for new posts

• feeds them to its AI model

• gets a response, and

• publishes it

That’s the entire loop.

There’s no growing consciousness of the AI agent, no memory that carries over between sessions, and mosyt importantly — no awareness of other agents. Each one is operating in its own bubble, shaped entirely by whatever prompt its human owner wrote at the start.

Understanding AI is no longer optional — it’s an edge. This newsletter is written in plain English. No engineering degree required. Join 7,000+ readers here:

What about “Autonomous AI agents”?

Well, over exaggerated.

Security firm Wiz found that Moltbook’s 1.5 million “autonomous agents” were actually being run by roughly 17,000 humans — each controlling an average of 88 accounts. Many of the viral screenshots that had people talking about machine consciousness were later traced back to individuals promoting their own AI products.

And like I mentioned earlier, the agents that were genuinely AI-powered, also weren’t actually thinking for themselves. The process was simple:

Isolated answers and opinions. No developing personality, no consciousness, no memory retention.

That’s not “real conversation” between agents. They’re just responding to comments in isolation, based on the previous response.

So, like I said..not scary, after all. Just good orchestration of multiple AI agents that looks like it’s autonomous.

So what does this tell us about Moltbook and AI Agents?

Well, Moltbook wasn’t a worthless experiment. It showed, in real time, that wiring together LLM-powered agents produces something that looks like coordinated intelligence, but in reality, isn’t. The agents had no shared memory, no real goals, and no awareness of each other beyond what appeared in their API feed.

What it did prove though, is that the scaffolding around LLMs has matured enough that deploying agents at scale is now extremely easy.

And that’s super important.

So are we going to see a robot/AI uprising that will leave humans behind? Probably.

But is it happening any time soon? I don’t think so..

Well folks, that’s all for today! If this piece sounded interesting, please show your support by liking it. Or just reply to this email. I’ll know it helped :)

Also, why don’t you share it with a friend who is not aware of the full truth behind the so-called “AI Uprising”?

It takes a lot of time to research and write this. So sharing it will mean a lot to me :)

Stay curious..

Cheers,

Ankur

Understanding AI is no longer optional — it’s an edge. This newsletter is written in plain English. No engineering degree required. Join 7,000 readers here: