The release valve

Make AI admit it doesn't know

One line that stops more hallucinations than anything else I do — and how to make it permanent.

Start here: paste this

Here from the video? This is the line. Add it to the end of any question where a made-up answer would cost you — anything involving facts, numbers, sources, or your own documents.

The release valve — add it to any prompt

If you don't know, that's fine, just say.

Make it permanent

Typing it every time works, but you'll forget on exactly the day it matters. Two ways to set it once:

In the Claude app (works in ChatGPT too)

Claude: Settings → General → the Instructions for Claude box. ChatGPT: Settings → Personalization → Custom instructions. Paste this and save — it applies to every chat from then on:

For your custom instructions

If you don't know something, or you can't verify it, say so plainly rather than guessing. "I don't know" and "I'm not certain, but here's my best understanding" are always acceptable answers here. Never invent facts, numbers, citations or quotes to fill a gap.

If you use Claude Code

Paste this into Claude Code inside your project — it writes the rule into your project's CLAUDE.md itself and shows you what it added. (Never heard of CLAUDE.md? Start with the guide.)

Paste this into Claude Code

Please add an honesty rule to this project's CLAUDE.md (create the file if it doesn't exist) so you stop guessing when you don't know something. The rule: if you don't know something or can't verify it from this project, say so plainly instead of producing something that sounds right. "I don't know" is a helpful answer here. Never invent facts, numbers, file contents, citations or API behaviour. Merge it with whatever is already in the file — don't overwrite or remove anything. When you're done, show me exactly what you added, and then tell me the most recent answer you gave in this project where this rule would have changed your reply.

Why it works

Think of coming home to find something broken and your kid standing suspiciously close to it. You ask if they did it, they say no, and suddenly there's an elaborate story involving the cat. Then you say “it's okay if you did” — and the truth comes straight out. You've made an environment where honesty is rewarded.

Models behave the same way. They're trained to be helpful and rewarded for giving you an answer — so when they don't know something, the path of least resistance is to produce something that sounds right. That's what a hallucination actually is: not lying exactly, just filling the silence.

The release valve works because it doesn't fight that training — it redirects it. You've made “I don't know” a helpful answer, so the model gets its reward without having to invent anything.

When it doesn't work

-It only removes the pressure to invent. If the model confidently believes something that's wrong or out of date, it won't say “I don't know” — it doesn't know it doesn't know. Important facts still get checked at the source.
-It catches a good chunk of hallucinations, not all of them. Nothing catches all of them.
-Don't use it to dodge judgement calls. On “should I do X or Y” questions you usually want its best reasoning, not an easy out — the valve is for facts, not opinions.

The other half of the problem

Inventing answers is one half of the same training problem. The other half is agreeing with you when you're wrong — same reward, different symptom. The fix for that one is a fuller instruction: Make your AI stop agreeing with you.

Want a hand putting this into practice? I do one-to-one AI coaching — wherever you're starting from.Learn more