I write this newsletter about not trusting AI on faith. Then I spent two days building a project with one, and watched my own argument play out on me, in real time, harder than I'd have written it.
The pattern was always the same. I'd ask for something. The AI would produce it, clean, confident, well-organized, the work of something that clearly knew what it was doing. And most of the time it was good. That's the part that matters. It wasn't bad at the task. It was good enough that trusting it felt safe.
Then I'd push. Is that everything? Did you check that? And a gap would open up. Not a dramatic one. A quietly missing piece the confident version had papered over. I'd point it out, the AI would fix it, agree completely, and hand back a new version that sounded just as finished as the last one. Which I'd also have to check. We did this for two days.
Here's what I want you to take from that, because it's the whole point of this newsletter and I got to live it instead of just argue it.
The failure was never that the AI was wrong in some obvious way. A model that's obviously wrong is easy, you catch it and move on. The failure was that it was confidently incomplete. It produced work that looked whole, sounded authoritative, and was missing things I only found because I refused to accept the first answer. The polish wasn't evidence of completeness. It was just polish. The two have nothing to do with each other, and that's the trap, because every instinct you have says a thing that sounds this finished probably is.
This is the part people adopting these systems are getting wrong. They're judging AI output the way they judge a person's: if it sounds competent and sure, it probably did the job. With a person that heuristic mostly works, because a competent-sounding person usually is one, and they feel the cost of being caught wrong. The AI feels nothing, and sounds exactly as sure either way. So the confidence tells you nothing, and if confidence is what you're using to decide whether to check, you won't check, and the gaps ship.
What actually worked was treating the AI like something that needed direction and verification at every step, not delegation. Concretely, the habits that earned good output:
Tell it to separate what it knows from what it's inferring or guessing, and to flag anything it can't verify, out loud, in the response. A model will produce a confident sentence and a wild guess in the same voice unless you force it to mark the difference.
Never accept "done" as done. "Is that everything?" found a gap almost every time I asked it. Not because the AI was lazy, but because "complete" and "looks complete" are different states and only one of them is visible in the output.
Assume the first answer is a draft, always. The good version usually showed up on the second or third pass, after I'd named what was thin. The first pass is what plausible looks like, not what correct looks like.
Make it do the verifying it can do. Cite the source. Check the current fact instead of recalling it. Show the work. The model can often verify itself if you require it, what it won't do is volunteer the verification unasked.
Now the uncomfortable question, because I'd be doing the exact thing I warn against if I skipped it. If an AI needs this much direction and re-checking, is it saving you anything? Honest answer: it didn't remove the work, it moved it. The AI did the producing. I did the deciding what was right and proving the rest. Which is the thing I keep saying in this newsletter, the verification is the work now, except this time I wasn't writing it about someone else's systems. I was doing it, for two days, on my own.
That's the whole lesson, and it's why I'm not worried about these tools the way some people are and not relaxed about them the way others are. They're genuinely capable. They also produce confident, complete-looking work with quiet holes in it, and they will hand it to you in a voice that invites you to skip the check. The value is real. So is the discipline it demands. The people who get the upside without the damage are the ones who treat a capable AI as exactly that: capable, and unverified, until they've done the part the machine can't, which is decide whether it's actually right.
I build a whole publication on that idea. It still took two days of doing it to feel how strong the pull is to just trust the confident answer. Resist it. That pull is the risk.
The prompts
A short field kit. Paste these into any AI to pull accuracy up and fabrication down. They won't make a model perfect. They make its guessing visible, which is the part you need.
Before you trust an answer
- "Separate what you know from what you're inferring from what you're guessing. Label each part."
- "Mark every claim you can't verify. If you're not sure, say so instead of filling the gap."
- "What would have to be true for this to be wrong? Check those things."
To catch fabrication
- "For every fact, statistic, name, date, or quote: give the source, or tell me you can't verify it. Do not invent sources."
- "Don't paraphrase a source so closely it's really a quote. If it's a quote, mark it."
- "If you don't know, say 'I don't know.' A short honest gap beats a confident paragraph."
To catch incompleteness
- "Is that everything? What did you leave out, and what did you assume?"
- "What's the weakest part of this answer? Argue against it."
- "List what you're NOT certain about in this response."
For current or factual things
- "Search and cite a current source instead of answering from memory. If you can't, tell me."
- "What's your basis for this, training data or something you just looked up?"
To force a second pass
- "Treat your first answer as a draft. Now critique it as if a skeptic wrote it, then give me the better version."
- "Where might you be pattern-matching to something common instead of answering my actual question?"
The one to internalize
- Confident is not correct. Complete-looking is not complete. Make it prove it.