Definition of Done Toolkit

Make your AI prove it's finished.

AI tells you "done!" when the work is wrong, half-finished, or quietly broke something else. This free kit fixes that. Takes about 3 minutes. No tech skills needed.

Works with ChatGPT Claude Claude Code Claude Cowork Codex any AI chat

What's a Definition of Done?

A short checklist that makes your AI prove its work is actually finished, before you trust it. Think of it as a receipt for the job.

  • What was the real problem?confirmed · figured out · or just guessing?
  • Who ELSE checked it?not the AI grading its own homework
  • How do you KNOW it works?something you can actually see
  • Where's the PROOF?a file or link that outlives the chat
  • How might it be WRONG?hunt the most likely failure first

No proof? Not done. ✎

The five things a good Definition of Done makes your AI answer.

Make your AI show you the receipts

Three easy steps, about 3 minutes. No tech skills needed.

  1. Copy the DOD Coach. Press the button below.
  2. Paste it into a fresh chat with your AI (or attach the file).
  3. Type start here and answer a few easy questions. It builds your Definition of Done.

Prefer to hand it over as a file? Download the coach, give it to your AI, and say "use this."

Want the full framework?

The coach keeps things simple on purpose. When you're ready to go deeper, there's a one-page reference with the full checklist for serious work: root-cause, review, verified, adversarial, and evidence.

Read the full Definition of Done reference →

Your AI is lying to you (and why this matters)

Your AI will tell you it's done when it isn't. It isn't trying to trick you. It reaches for "it's done!" because that's the answer that sounds helpful, so out it comes.

Then you go look. The work is wrong, or only half-there, or it quietly broke something else on the way.

And when you catch it? It folds instantly. "You're right to push back. I'm sorry. I made that up." Every single time. But by then you've already trusted it and moved on. That's the part that costs you.

Here's why it happens, plainly.

Every AI you'll touch works the same way underneath. ChatGPT, Claude, Gemini, all of them. It's a giant pattern-matcher.

It reads your words and predicts the words that most likely come next. That's the whole trick. On its own it isn't looking anything up or checking against the real world, it's just producing the answer that sounds right. (Some tools can now search the web or run code, which helps a lot. The thing underneath is still guessing.)

Most of the time, "sounds right" and "is right" line up. That's why it feels like magic.

But when they don't, the AI has no idea. It will invent a fact, a quote, a source, a function that doesn't exist, a test result that never ran, and hand it to you dead certain. People call this "hallucinating."

The same machinery that makes the AI useful is the machinery that hallucinates. It's one system, running on a question where the pattern and the truth happen to disagree, so you can't file it as a bug and wait for a fix.

And here's the uncomfortable bit: you can't fully turn it off. There's no setting for it, and no prompt or newer model has made it go away. Anyone who tells you their AI "doesn't hallucinate" is selling you something.

So most people reach for the obvious fix. They tell it to stop. "Don't make things up." "Only tell me if you're sure." They pile on longer prompts and sterner warnings.

It mostly doesn't work. You're asking a pattern-matcher to promise it'll be accurate, and it will happily promise, in the same confident voice it uses for everything, then carry on exactly as before. A vague rule like "be accurate" is just more words for it to predict against.

The shift that actually works is smaller, and way more powerful. Stop piling on vague rules and give it a clear Definition of Done instead: concrete things it has to hand you before it's allowed to call the job finished.

Think of it like a receipt. You wouldn't let a contractor walk out of your house without walking the job first. Same energy. Before you accept the work, make the AI walk you through the proof instead of just claiming it's there.

That's the whole job of a Definition of Done. It comes down to a handful of habits:

And this matters just as much for knowledge work as it does for coding.

A coder needs it because the AI will confidently call a function nobody wrote and swear the tests passed. But a writer, a researcher, or anyone running a business needs it every bit as much. The AI will invent a statistic, misquote a source, or summarize a document it only half-read, and all of it lands polished and sure of itself.

That polish is the whole trap. A Definition of Done is how you stop trusting the confidence and start checking the work.

Don't overthink it. If you take one thing from this page, take this question and start asking it:

"Before you tell me you're done, show me one thing I can check to know it actually worked."

If it can't show you, it isn't done. That one line already puts you ahead of almost everybody.

What "done" actually asks the AI to prove

When you get serious, a good Definition of Done makes the AI answer a few questions before it's allowed to say it's finished. In plain English:

You don't have to memorize any of this. The coach turns it into a short checklist in your own words, shaped around how you actually work. Writing, coding, research, running a business, making decisions. Whatever your thing is.

Where people get stuck

Four snags trip up almost everyone learning this. If one of them is you, good. That's the whole point.

1. "I don't even know what done looks like for my task."

This is the most common one, and it's the deepest. If you can't say what finished looks like, the AI will happily invent its own version and call it done. But you don't have to figure it out alone. The coach asks you a few plain questions about your work and your goal, then helps you name it. And if you still can't name it? That's a signal the task isn't ready to start. Good thing to know before you burn an hour on it.

2. "Is it telling me the truth, or just making it up?"

You can't always tell. So make the AI tell you. Have it label every claim: confirmed (real proof, like an error message or a test), figured out (it reasoned it, but has no proof), or a guess. Now you know where to push. "Confirmed"? Ask it to show you the proof. "Guess"? Don't build on it yet.

3. "Who's supposed to check it? It's just me in a chat."

People assume you need fancy tools for a "second set of eyes." You don't. The AI is bad at catching its own mistakes, same as us. So the trick is a genuinely fresh read. Paste the finished work into a new chat, even a different AI, and say: "read this like you've never seen it. What's unclear? What could go wrong? What did it miss?" Two separate reads catch what one glosses over.

4. "Isn't this just planning?"

Nope. They're opposite ends of the same task, and you want both. Planning is the front: what am I doing, and what are the steps? Definition of done is the back: did I actually do it, and how do I know? One more thing. The proof has to outlive the chat. If the only record is in a conversation that vanishes tomorrow, it doesn't count. Good gut-check: "could someone read this in six months and know what got done, and how we know it worked?"