ChatGPT Productivity

Prompts for Checking an AI's Technical Work

MedChemBench found a model scoring 95% at reading molecular structures and about half at deciding which one to test next. OpenAI's chip work ran every model rewrite through a formal equivalence checker before anyone trusted it. The pattern holds generally: AI is strong at reading a specification and weak at the judgement call, so the useful prompts are the ones that separate the two and make the model show which it is doing. These six do that.

⌨️ 6 prompts 🕐 Updated Nov 22, 2026 👁 1 views
💡 How to use these prompts: Replace everything in [BRACKETS] with your specific details before sending. Click Copy to copy any prompt to your clipboard instantly.
1
Separate what was read from what was inferred
Verification
Answer this, then split your answer into two labelled lists.

READ: facts you took directly from the material I gave you, each with the line or section it came from.

INFERRED: anything you concluded, assumed, or filled in from your own knowledge.

If the INFERRED list is longer than the READ list, say so at the top before anything else.

[paste the material]
2
Make it rank its own confidence
Verification
Give me your answer, then a table with one row per claim you made.

Columns: the claim, how confident you are from 1 to 5, and what would prove it wrong.

Any claim you rate 3 or below, rewrite it as a question I should check rather than a statement you are asserting.

Do not pad the list - a claim I would not act on does not need a row.
3
Force the equivalence check
Code review
Rewrite [this function / query / config] to be [faster / clearer / smaller].

Then prove the behaviour is unchanged: list every input class the original handles, and what each one returns before and after. Include the edge cases - empty, null, maximum size, wrong type.

If any input behaves differently after your rewrite, stop and tell me instead of explaining why it is fine.
4
Ask what it would need to decide properly
Decisions
I need to choose between [options]. Before recommending anything, tell me what information you are missing that would change the answer.

List it as: the missing fact, which option it would favour, and how I could find it.

Only after that, give me your recommendation on what you currently know, and say which of the missing facts would most change it.
5
Check a number before you quote it
Research
You stated [figure]. Before I use it anywhere:

Where exactly did it come from - which source, which date, measured how?
What is it relative to? A percentage with no baseline is not a number.
What has changed since it was published that could move it?

If you cannot answer all three, tell me the figure is unverified rather than restating it.
6
Make it argue the other side
Decisions
You recommended [X]. Now make the strongest case against it that an informed sceptic would make.

Not a list of generic risks - the specific reason this is the wrong call for my situation, using what I have told you.

Then tell me which of those objections you cannot answer, and whether it is enough to change your recommendation.