What makes a good AI prompt

A good prompt tells the model what a good answer looks like before it starts. Most weak prompts are not wrong, just underspecified: they leave the model to guess who the answer is for, how long it should be, and what shape it should take. The model then fills the gaps with something generic.

The six parts most prompts need

1. Role: who the model should be, or who it writes for. "You are a patient tutor" or "Write for busy managers".
2. Context: the material and background it cannot know: your notes, your product, your audience.
3. Task: one clear instruction with a verb. Summarize, rewrite, compare, draft, list.
4. Format: the shape of the answer. Headings, a table, bullets, JSON with given keys.
5. Constraints: length, tone, what to avoid, and what to do when information is missing.
6. Examples: one short sample of the output you want. It often does more than another paragraph of instructions.

A quick question does not need all six. A prompt you reuse every week usually does.

A complete example

Here five of the six parts are separate sections (this one needs no example), and the notes that change each week are a variable:

<ROLE>
You are an editor for a busy product team.
</ROLE>

<CONTEXT>
The notes below are from our weekly planning call:
{{notes}}
</CONTEXT>

<TASK>
Summarize the decisions and the open questions.
</TASK>

<FORMAT>
Two headings, Decisions and Open questions, with at most five bullets each.
</FORMAT>

<CONSTRAINTS>
Use only what the notes say. If something is unclear, list it under Open questions.
</CONSTRAINTS>

Compare it with "Summarize these notes": the reader, the length, the structure and the rule for unclear points are now stated rather than guessed.

Make it reusable

Put the parts that change in {{double_braces}} and keep the rest fixed. Read how to use variables in AI prompt templates, or try the template filler to fill one in.

Test it, then change one thing at a time

• Run it on two or three real inputs, including an awkward one: empty notes, a very long document, a mixed language.
• When an answer misses, change one instruction and run the same inputs again.
• Keep the earlier version so you can compare. The prompt diff shows exactly what changed.

Paste your own prompt into the prompt checker to see whether it states a role or audience, a format, a length and an example, or let the prompt refiner suggest a clearer version.