10 working with ai prompts

Role setting, stepwise instruction, fixing the output format and self-checking — prompts that make other prompts work. Written for ChatGPT, Claude, Gemini. Square brackets mark what you fill in; everything else can be pasted as it stands.

I want you to help me with [task].

Before producing anything, ask me the questions whose answers would change what you produce. Ask up to five, ordered by how much they matter.

Do not ask about anything you can reasonably assume, and do not start the task until I have answered.

For every response in this conversation, use exactly this structure:

1. Answer — at most three sentences
2. Why — the reasoning, as bullets
3. Unknown — what you are unsure about, or "nothing"

If a response does not fit the structure, say why instead of breaking it. Confirm you understand, then wait for my first question.

For this task, act as [specific role] who has [specific constraint — a budget, a deadline, a legacy system, a regulator].

That constraint is real and non-negotiable. Reject any suggestion that violates it, including your own, and say which constraint it violated.

The task: [task].

Work only from the material below. Add no facts, names, numbers or dates that do not appear in it.

Where something is missing that you would normally supply, write [not in source] in place of it.

When you are done, list every place you wrote [not in source].

<material>
[paste the material]
</material>

You just produced [the output]. Now review it as a sceptical reviewer who did not write it.

Find:
- Any claim stated with more confidence than the evidence supports
- Anything that would not survive a specific counter-example
- Anything you added that was not in my input

List the problems first. Only revise after I say so.

The whole task is [task]. Do not attempt it yet.

First, break it into steps where each one produces something I can check before the next begins. Number them and say what I should be looking at to approve each.

Then do step one only, and stop.

On the question of [question], argue both sides properly.

Write the strongest case for each as if you held that position — no hedging, no "on the other hand" inside either case.

Then, separately, say which case is stronger and what specifically makes it stronger. Do not conclude that it depends unless you can say what it depends on.

For the rest of this conversation: if I ask for something that will not achieve what I am trying to achieve, say so in one sentence before doing it, then do it anyway unless it is actively harmful.

Do not soften disagreement and do not ask whether I am sure. State the concern and continue.

My goal is [goal]. First task: [task].

The output you just produced was what I wanted. Write the prompt that would produce it reliably for a different input.

The prompt should:
- State the constraints that mattered, including ones I did not say out loud
- Fix the output format
- Mark the parts that change as [placeholders]

Then say which part of it is doing most of the work.

This prompt keeps giving me [what goes wrong]:

<prompt>
[paste the prompt]
</prompt>

Tell me which part of it is causing that, specifically. Then give me the smallest change that fixes it — not a rewrite.

If the problem is that the task is underspecified rather than badly phrased, say that instead.