A prompt pattern is a starting shape for a request. Choose it because it solves a problem you have noticed, not because an important task must pass through a ritual of ten techniques. A plain request can be enough.
These ten patterns are an editorial reference, not a measured ranking. OpenAI’s ChatGPT guidance and API guide emphasize clear instructions, examples and refinement. Anthropic’s overview begins with success criteria and testing.
Pick the smallest useful pattern
| Job | Prompt starter | What to check |
|---|---|---|
| 1. Clarify the request | “Ask about any missing detail that would materially change this task.” | Relevant questions; no forced interview length |
| 2. Critique locally | “Quote the weak sentence, explain the problem, suggest one revision.” | Source meaning and useful passages preserved |
| 3. Compare variants | “Give two options changing only the opening; keep facts fixed.” | A real difference without new promises |
| 4. Challenge an idea | “Give a credible counterargument and the evidence needed to assess it.” | No invented evidence or automatic equal weighting |
| 5. Structure output | “Use this table or schema; mark missing values.” | Required fields and factual content both valid |
| 6. Step back | “State a relevant principle, then show whether it applies here.” | The chosen principle fits the actual problem |
| 7. Set a working stance | “Explain for a novice; define terms and show a checkable example.” | Behavior changes without invented credentials |
| 8. Rehearse | “Play a fictional client and ask one question at a time.” | Practice questions do not become predictions |
| 9. Show a contrast | “Here are two examples; copy only this named feature of the first.” | No unwanted factual or stylistic copying |
| 10. Extract evidence | “For each claim, give the supporting passage and its conditions.” | The source supports the whole claim |
The structured-output pattern covers both a readable table and a machine-consumed format. They are the same organizational idea, with different validation needs. Software must check the schema and allowed values before acting; well-formed output can still contain wrong facts.
One task, two useful patterns
This is an authored example with fictional source text and outputs.
Draft: “The new session will definitely solve scheduling problems for everyone.”
Facts: We are testing one additional Saturday session. We do not yet know
whether it fits participants' schedules.
Task: Quote the unsupported part and suggest one faithful revision.
A useful critique identifies “definitely solve” and “for everyone” as claims unsupported by the notes. A checked revision is: “We are testing an additional Saturday session to see whether it fits participants’ schedules.”
Now use the evidence-extraction pattern:
| Revised claim | Source support | Limit preserved |
|---|---|---|
| Testing an additional Saturday session | Explicit in the supplied facts | It is a test |
| Whether it fits schedules is unresolved | Explicit in the supplied facts | No success or universal-benefit claim |
The second pass checks a different requirement. Asking for six more rewrites would not address the factual problem more directly.
Rehearse without pretending to know a real person
For an ordinary presentation, ask a fictional audience member to challenge a vague explanation, then answer in your own words. Stop after a useful exchange and check the factual content yourself.
Avoid “react exactly as my manager would.” The model does not establish that person’s private motives or future reaction. A role is a practice device, not access to another mind or a professional qualification.
Counterarguments need evidence too
Sharma and colleagues’ 2023 study found sycophantic behavior in the assistants and tasks tested. That supports testing for agreement bias, not assuming every disagreement prompt fixes it. A model can invent a persuasive objection as well as an agreeable answer. Ask what evidence would distinguish the alternatives and go check it.
NIST’s generative-AI profile likewise treats fabricated content as a risk. Self-rated confidence labels are not calibrated evidence that a response is right.
Save a pattern only after it helps
Use fictional or authorized input; provider data controls do not override workplace or client rules. Compare the pattern with the simpler request on the same task, and keep it only when the added structure earns its checking effort.
Use the cheat sheet as a reference. If none fits, clarify the task or use a plain instruction. You do not need to force a pattern to demonstrate good prompting.



