A prompt is an interface contract
A useful prompt states the objective, relevant context, source material, constraints, output format and quality criteria. If any of those are ambiguous, the model must guess. Better prompts reduce unnecessary guessing rather than trying to “unlock” hidden intelligence.
Separate evidence from instructions
Place source material in a clearly delimited section and tell the model what it may infer beyond that evidence. When citations matter, require source references for factual claims and allow the model to say that the evidence is insufficient.
Design the output before the wording
If another person or system will consume the result, define the structure first: a table, JSON schema, checklist, decision memo or diff. A stable output contract makes evaluation and automation much easier.
Use examples when the criterion is subtle
Few-shot examples are valuable when tone or labeling rules are difficult to explain. Include both a good example and, when useful, a counterexample. Avoid examples that accidentally teach irrelevant patterns.
Build a verification pass
For important work, ask a second step to check missing evidence, contradictions, unsupported claims and format violations. Better still, validate mechanically where possible and use a small eval set to detect regressions over time.