1. Define the action and purpose
Start with the action—summarize, compare, draft, review, or convert—and the outcome. 'Extract decisions and owner-specific actions from these meeting notes' is clearer than 'look at this.'
When one request contains unrelated jobs, give priorities or split it into stages.
2. Provide relevant context and source
State the audience, use case, current situation, and reference text. Define internal terms instead of asking the model to guess them.
Do not submit personal data, passwords, API keys, or confidential company material to an external AI. Replace identifiers with safe examples when possible.
- Audience and purpose
- Background and terminology
- Only material you may disclose
3. Specify format and constraints
Set length, language, tone, and a format such as a table, list, or JSON. Name required and excluded content.
For example: 'Write in Korean, under 500 characters, use an improvement/benefit table, and label assumptions.'
4. Request evidence; 5. review the result
For accuracy-sensitive tasks, ask the model to avoid guessing, mark uncertainty, and point to the supplied basis.
A person must still check facts, numbers, links, code, and privacy. A stronger prompt improves the draft but does not transfer final responsibility to AI.