Reading Time: 8 minutes

Hey Prompt Lover,

Yesterday I asked whether AI was making people less capable.

Doug replied in four words. "No. I do way more now than I ever did before."

That was it. No elaboration. Just a clean pushback and a name.

And Doug is not wrong. Which is exactly why yesterday's question needed a follow-up.

Jeff Pogue came back with a second email and he did something no one else in this series has done yet. He flipped the question entirely.

He told me about Nanobanana. An AI image tool. He said he has never been able to design. Never had that skill. In the past he would have hired a designer, tried to describe what was in his head, gone back and forth through revisions, and eventually landed somewhere close to what he wanted.

Now he chats with an image tool and gets there in one afternoon.

Then he told me about Lovable. He was good with Wix after years of practice. His new website built with Lovable is better in every way. Features he always wanted but couldn't get from Wix templates. One afternoon. Done.

Then he asked me something;

"What about the opposite of the question? Did AI suddenly give us a capability that we did not have before? At the very least, a capability that we can now do with our own agency and not require a team?"

That is the question nobody is asking in the capability conversation.

ChatGPT is a superpower if you know how to use it correctly.

Discover how HubSpot's guide to AI can elevate both your productivity and creativity to get more things done.

Learn to automate tasks, enhance decision-making, and foster innovation with the power of AI.

Download the free guide

For three days we have been talking about what AI takes. The skills that atrophy. The understanding that doesn't develop. The competence debt that accumulates quietly.

Jeff is pointing at something real on the other side of that.

There are things people could not do before AI that they can now do alone. Design.

Build websites. Write code. Produce music. Create legal documents. Build financial models. Run data analysis. Produce video content.

For most of human history, access to these capabilities required one of two things. You either spent years developing the skill yourself or you hired someone who had. Both paths had real friction. Real cost. Real time.

AI collapsed both of those paths for a lot of people simultaneously.

Jeff building his own website with features he always wanted but couldn't build — that is not a skill he had and lost. That is a capability he never had and now has. The wow he feels when he sees the formula is not the feeling of losing something. It is the feeling of access to something that was always behind a wall that has now been removed.

And here is where it gets complicated.

Is that liberation or is that borrowing?

The CFO question he raised at the end of his email is the sharpest version of this.

"Can AI effectively replace that role and save your company head count? Should it?"

Think about the arc of that question carefully.

You could not replace your CFO before AI. The knowledge required was too specific. The judgment too contextual. The accountability too real.

Now AI can produce the analysis, the models, the forecasting, the reporting that a CFO would produce. Not perfectly. Not in every situation. But well enough that the question Jeff is asking is genuine and is being asked in real boardrooms right now.

If you can access CFO-level financial analysis through an AI tool without hiring a CFO, have you gained a capability or have you borrowed one without the understanding to know when it's wrong?

And when it is wrong — when the model is confident about something that is incorrect — will you know?

Ask Attio anything: prep for calls, update records, spot deals at risk Over 4,000 startups including Granola, Wispr Flow, and Lightdash, already use Attio. Try Attio for free.

Have you ever fact-checked something Claude told you and found it was wrong? What did you do next?

This follows directly from the capability conversation.

Because the new capability Jeff is describing comes with a hidden condition. You can now do things that used to require experts. You can access outputs that used to require years of training. You can produce things that would have been impossible alone.

But the tool that gives you that access has a specific and well-documented failure mode. It produces confident, fluent, plausible output that is sometimes simply wrong. Not vague. Not uncertain. Wrong in a way that sounds exactly like right.

And the less you understand the domain, the harder it is to catch.

Jeff building a website he couldn't build before — he can look at the result and know if it works. He can click every button. He can see what loads and what doesn't. The feedback loop is immediate and visible.

Jeff's AI-generated financial model — the feedback loop is invisible until the business makes a decision based on numbers that were wrong. By then the cost of being wrong is much higher than the cost of the website not loading.

This is the part of the capability conversation that almost nobody is having.

Before I get to my honest answer today, something happened in the inbox yesterday that I have to mention.

Owen wrote in. He appreciated the content but asked a direct question.

"Curious if anyone ever mentioned how ChatGPT'd the writing style is?"

Then he added that he publishes a newsletter too and has been trying to solve the same problem.

I replied to Owen and told him what works for me. Going into prompts being very specific about not sounding like AI. Catching the telltale patterns before they leave. Reading back aloud and fixing anything that sounds like it came from a machine.

But Owen's observation sits inside this series in a way that I cannot ignore.

We are four days into a series asking uncomfortable questions about AI and human identity. And a reader is pointing out that the newsletter itself sounds AI-generated.

I am not going to pretend that is not an uncomfortable thing to notice. It is exactly the kind of thing this series is here to surface.

The tool I am using to write about AI changing writing is changing how I write.

That is either the most honest thing this series has produced or the most ironic. Probably both.

The last time Claude gave me something confidently wrong was a statistic I used in a research piece. I quoted it. I sent it. It was wrong. I found out later.

What I did next was add the verification step I described in the prompting newsletter a few weeks ago. Ask Claude to rate its own confidence on every factual claim before it goes anywhere.

What I should have done first was verify it myself before trusting a tool's confidence at face value.

The capability conversation and the hallucination conversation are the same conversation.

You gain access to things you couldn't do before. You lose some of the instincts that used to protect you in the domains you do know. You use the tool in domains you don't know. The tool is confident. You don't have the knowledge to check it. And something gets through that shouldn't.

That is the gap nobody is talking about when they talk about AI giving you new capabilities.

Today's question for the inbox.

Think of the last time you used AI for something you don't fully understand yourself. Did you verify the output before acting on it? What would have happened if you hadn't?

Reply and tell me. And if you caught something wrong — what was it and what did you do?

Day 5 arrives tomorrow.

— Prompt Guy

Day 4 of 30. One question every day that the AI industry is not asking. If you know someone who should be in this conversation, forward this to them now.