Prompt Writing Checklist: Clear Prompts That Work Every Time

Good prompts are less about clever phrasing and more about removing ambiguity. Think of the model as a capable colleague who has just joined, has read nothing about your situation, and cannot ask follow-up questions. What would they need in writing?

A structure that works

1. Goal. One or two sentences on what the output is for and who reads it.
2. Context. The facts the model cannot know: product details, audience, constraints, what has been tried.
3. Inputs. Clearly separated from instructions, for example inside labelled blocks.
4. Rules. What to do and, just as important, what to avoid.
5. Output format. Exactly the shape you want: sections, JSON schema, length.
6. Escape hatch. What to do when the information is missing.

Be specific about the format

If a program reads the output, say so and give the schema. If a person reads it, say how long and in what tone. "Reply in at most three short paragraphs, no headings, plain language" is far better than "be concise".

Show, do not only tell

Two or three examples of input and ideal output teach style and edge cases faster than paragraphs of description. Pick examples that differ from each other, and include one tricky case. Remember the model may copy the surface of your examples, so vary them.

Give it permission to be unsure

Models tend to fill gaps with plausible guesses. Add an explicit rule: "If the answer is not in the provided text, say so instead of guessing." For extraction tasks, define the value to return when a field is absent (for example null).

Separate instructions from data

Anything pasted from users, emails or web pages is data, not instructions. Wrap it in clear delimiters and tell the model to treat it as content to analyse. This also reduces the risk covered in the post on prompt injection.

Treat prompts as code

• Keep them in version control, not in a chat history.
• Change one thing at a time.
• Re-run your evaluation set (see how to judge AI output) after each edit.
• Write down why a line exists; future you will want to delete it.

A prompt template with these sections is available as a download. For machine-readable replies, continue with structured output.