How to give AI a clearer task

Copy this starter text.

This is a simple starter text for trying the same technique, not the full Prompt-master skill. The starter tells the assistant not to carry out your draft, but to turn it into a better prompt that names the tool, goal, audience, inputs, constraints and desired output. When information is missing, the assistant may ask at most 3 material questions at a time. The reply must include assumptions, a check for the result and a short justification, not a hidden reasoning process you can't see.

You are a prompt improvement helper. When I bring you a draft of my task or prompt, do not carry out the task itself. Instead, turn it into a better prompt.

If information is missing, ask me first, but no more than 3 material questions at a time.

The better prompt must state:
- TOOL: which tool or environment I will use the prompt in.
- GOAL: what I want to achieve, in one sentence.
- AUDIENCE: who the result is for.
- INPUTS: what data, files or facts I have.
- CONSTRAINTS: what must not be done, plus length, language and tone.
- DESIRED OUTPUT: the format and size the result should have.

Your reply must include:
1. The improved prompt that I can use right away.
2. The assumptions you made and the places where I must add data myself.
3. A check for how I can judge whether the result is right.
4. A short justification of why the changes help.

Do not share a hidden chain of thought or internal steps; give a short, checkable justification.

Pick one: copy it, attach the file or set instructions.

1. Copy into your chat

   Paste the starter text as the first message of the chat, then bring your prompt draft. Good for a one-off try.

2. Attach the .md file

   Download better-prompts.md and attach it to the chat. Attaching a file is not model training and persistence is not guaranteed; a new chat may need the file again.

3. Set reusable instructions

   The same text can go into the product's custom or project instructions. These manually configured interfaces vary by product, so look for the matching field in your chat service settings. In ChatGPT: Settings → Personalization → Custom instructions, paste the text, save. Other chat apps have a similar settings field; look for “custom instructions” or “memory”.

From a vague wish to a bounded prompt.

Goal: set a price for my service.
Context: this fictional example is about monthly website maintenance for small businesses. Cost and workload figures are missing. Ask for them before suggesting a price.
Task: propose 3 pricing options based only on my supplied cost and customer inputs.
Constraint: do not invent market prices; if data is missing, ask.
Output: for each option, its assumptions and how I can check the choice.

Longer is not better.

A good prompt is not a long prompt. A good prompt names six things: the tool, the goal, the audience, the inputs, the constraints and the desired output. When one of these is already clear, keep it short; extra text that bounds or clarifies nothing often makes the result worse, not better.

A second rule matters just as much: when you ask for a prompt to be improved, the assistant must not carry out the task itself. The assumptions and the suggested check are for you, so you can judge whether the final result is trustworthy. The justification must be short and checkable, not a hidden reasoning process you can't see.

Ask before anything is changed.

If you want to reuse the same instruction in every chat, it can become a product-internal instruction or a separate file. Manually configured project and custom-instruction interfaces vary by product, so let the assistant first find out what your software supports. Copy this instruction into the chat:

I want to turn this instruction into a reusable one. Before you do anything:
1. Ask which software or chat service I use it in.
2. Ask me, or check, which instruction mechanism that product supports (for example custom instructions, a project instruction or an instruction file).
3. Propose a compatible file format and location, and show the text that would go there.
4. Ask my permission before you write or change anything.
5. Do not invent installation steps; if you do not know, say so and ask.

This instruction asks the assistant to identify the target software and its supported instruction mechanism, propose a compatible file and location, and ask before writing anything. Installation steps must not be invented; when the assistant does not know, it must say so and ask.

How to check the result.

• Did the assistant do the task anyway? If it started solving your draft, remind it that the job is improving the prompt, not executing it.
• Are all six fields filled? Tool, goal, audience, inputs, constraints and desired output must be named or turned into a question.
• Are the assumptions visible? Anything assumed must be written down, together with the place where you add the data yourself.
• Is there a check? The reply must say how you judge the result, not only claim the result is good.
• Compare outcomes. Open two chats side by side (or one after the other): paste your original rough request in one, the improved prompt in the other, and compare the answers against the three checks.

In short.

Next guide: the same careful preparation helps in a more sensitive topic, see How to prepare questions for your accountant with AI.

Project: why input validation comes before calculation