Seven Prompt Patterns Worth Reusing

Prompt engineering is less a craft of clever phrasing than a small library of patterns, each fixing a specific failure. Learn the failure and the pattern becomes obvious.

1. Persona with expertise, not personality

"You are a compliance reviewer for EU consumer contracts" narrows vocabulary and priorities. "You are a friendly helpful assistant" narrows nothing. Fixes: answers pitched at the wrong depth.

2. Rubric in the prompt

State the criteria a good answer meets before asking for the answer. Fixes: output that is fine but not what you meant by good.

3. Few-shot for shape

Two or three examples teach format and tone far faster than description. Make one an edge case so the pattern generalises. Fixes: inconsistent structure across runs.

4. Decomposition

Split a compound request into ordered subtasks, each with its own output. Fixes: the model doing the first half thoroughly and the rest in one sentence.

5. Self-critique pass

Draft the answer.
Then list every requirement from the request and mark it met or missed.
Revise anything marked missed, and return only the revised answer.

Fixes: near-misses on explicit constraints. Costs an extra generation, so reserve it for high-stakes output.

6. Escape hatch

Name what to do when the model cannot comply: ask a question, return a specific null value, escalate. Fixes: confident invention when information is missing — the most damaging failure of all.

7. Format lock

Specify the exact output shape and forbid preamble. Fixes: "Sure! Here's your JSON:" breaking your parser at 3am.

Combining them

• Classification: persona + few-shot + format lock.
• Long-form generation: rubric + decomposition + escape hatch.
• Extraction: format lock + escape hatch + deterministic validation.
• High-stakes review: rubric + self-critique + human sign-off.

Reach for the pattern that matches your failure, not the one that sounds most sophisticated.