Prompt structure
| Prompt part |
Purpose |
| Task |
State the work to perform |
| Context |
Provide the audience, source material, scope, and constraints |
| Output format |
Request a table, checklist, summary, draft, code, or another usable form |
| Quality criteria |
Define accuracy, tone, coverage, citations, or other success conditions |
| Safety boundary |
Exclude sensitive data, unsupported claims, or unapproved actions |
1Summarize the attached approved policy for a support-team audience.
2Use five bullets, preserve exceptions, and identify claims that need source review.
3Do not add policy requirements that are absent from the source.
Verify before use
| Output type |
Appropriate check |
| Facts, numbers, or current claims |
Verify with a primary or authoritative source |
| Summary of supplied material |
Compare with the original for omissions and changed meaning |
| Code, formula, or configuration |
Test in a safe environment and review security implications |
| Recommendation |
Examine assumptions, alternatives, risks, and failure modes |
| High-impact decision support |
Require accountable human review before acting |
Sensitive-data boundary
Do not submit personal, confidential, regulated, credential, secret, proprietary, or production-sensitive data to an AI service unless the organization explicitly permits that data and use case in that service.
| Data |
Safer practice |
| Customer or employee records |
Use approved tools and minimize or anonymize the data |
| Credentials, keys, and tokens |
Never submit them; rotate any secret that is exposed |
| Production logs |
Redact tokens, identifiers, account data, and internal URLs |
| Legal, medical, or HR material |
Follow policy and require suitable human review |
Agent boundaries
| Capability |
Boundary |
| Read data |
Limit sources and verify retrieval quality |
| Edit content |
Review before publishing or sending |
| Call tools |
Grant least privilege and scope each tool to the task |
| Send messages or make changes |
Require approval for consequential actions and keep an audit trail |
| Resolve ambiguity |
Ask for clarification or stop safely rather than guessing |
An AI assistant can produce a fluent answer without establishing that it is true, permitted, or suitable. Treat confidence as a presentation quality, not proof.