When the AI says it's done

"The first principle is that you must not fool yourself — and you are the easiest person to fool." — Richard Feynman

I build a lot of things with AI now — small tools, a whole website, the works. And the more I do it, the more one habit keeps saving me: I don't take "done" at face value. Not from the AI, and not from myself. Here's what that looked like this week.

"Done" is a claim, not a fact

When the AI finishes something, it tells me it's done — and that it checked its own work. I've learned the hard way that those are two different things. So now I bring in a second AI to check the first one's work: a fresh one that never saw how the thing was built, and has no reason to be polite about it. Every single time, it has found something genuinely broken that the first one swore was fine. "Done," from an AI, is a claim — not a fact. Treat it that way.

The same lesson turned up in a different outfit. I took apart one of those popular little add-ons that promise to make your AI smarter — and instead of just installing it, I read what it actually does, and I read its own evidence honestly. It turns out the better your AI already is, the less the add-on does anything at all. The maker wasn't lying; he was perfectly fair. It's just that a confident pitch and real proof are two very different things. Read the receipts.

First check what you already own

One more that saved me a whole afternoon: "we need this on both sites" does not mean building it twice. Before I built a way for people to follow the site, I checked the other one — and it had quietly had that feature the whole time. Half the job was done before I started. Always look at what you already own before you build something new.

The tidy summary that hid the risky line

The shortcut I use to skim a web page hands me a tidy little summary. Genuinely useful — until you're trying to check what a stranger's tool would actually do before you let it near your computer. Then a tidy summary is the last thing you want: it smooths right over the one line that matters. The fix is boring, and it works — read the real thing, word for word, not the recap.

Prompt to steal: make a fresh AI check the work

This is that second opinion, boiled down to something you can paste. Point it at anything an AI just built for you:

✦ Prompt Claude · Code Copy

Don't trust that this is done — check it. Treat everything I told you as a claim, not a fact. Go and test each thing yourself instead of just reading it back to me. Look hard for anything quietly made weaker, anything marked done that isn't, corners cut, and leftover mess. Then give me a straight verdict — works, works with caveats, or doesn't — and show me the proof.

It works because it stops the AI defending its answer and gets it testing its answer instead — the closest a non-coder gets to a proper second opinion.

Fun project idea: vet a tool before you install it

A little helper that checks a stranger's AI add-on before you install it. It reads what the thing would actually do on your computer, and warns you if its instructions quietly try to make your AI act behind your back — the sneaky modern version of a dodgy download. Not a magic "safe or unsafe" button, because those just lull you into trusting. It shows you what it found, and leaves the decision to you.

Current project: little tools that show, not tell

• I've been shipping small interactive tools that make one tricky idea click by showing it instead of explaining it.
• My favorite so far takes a flat picture you simply cannot solve, and folds it up into three dimensions until the answer is suddenly, obviously there — no math required.
• And yes: that suspicious second AI went through every one of them and found a real bug in each, before it ever went live. Have a look at the CS Lab.

"Done" is a claim. Check it. ✦

If this saves you one moment of trusting a shiny "done," it did its job. See you next Sunday.