A good prompt is not a magic spell; it is a clear task brief. Break an unclear request into a goal, context, requirements, and checks to get a more useful first draft.
A reusable structure
- Goal: What result do I need?
- Background: What material is necessary, and what information may be used?
- Constraints: Specify audience, tone, length, format, and prohibitions.
- Check: Ask the model to flag uncertainty rather than invent sources.
Example: organize meeting notes
Turn the meeting notes below into minutes.
Audience: project members who did not attend.
Format: decisions, action items, owners, and due dates.
Use only information in the source; write “to be confirmed” when an owner or date is missing.
Meeting notes: [paste de-identified notes]Compared with “Write meeting minutes,” this version makes omissions easier to check. A person should still compare the result with the original notes.
Make the result more useful
Provide one or two desired output examples, without exposing sensitive data. For a complex task, split it into extraction, organization, and polishing. If citations are required, provide verifiable source text and check every citation.
Do not ignore data boundaries
Without authorization, do not upload customer information, credentials, or trade secrets. Review a third-party service’s data-processing terms and de-identify inputs when necessary.
Break up an “everything at once” request
Extract facts first, check missing information next, and only then write for a particular audience. Each step can be compared with the source, making errors easier to locate. Asking for ten deliverables at once often increases review costs.
Example: structured extraction
Extract the title, time, and location from the notice below.
Output only JSON, with fields named title, time, and location.
Use null for a value that does not appear; do not guess.
Notice: The project retrospective will take place in the third-floor meeting room on Friday.“Friday” is still a relative date and cannot become a calendar date without a reference date. Code should validate that the JSON parses and field types are correct; important dates also need human confirmation.
What if the answer is wrong?
Point out the specific error and add the missing evidence; ask for only the relevant part to be revised. If repeated attempts remain wrong, change the task or approach rather than endlessly adding dramatic role instructions. Automation also needs code validation and human review.