ChatGPT prompt patterns cheatsheet (the 10 patterns that ship)
Ten production-tested ChatGPT prompt patterns: role + objective, structured output, persona conditioning, step-back, output gating, anti-template, refusal calibration, evaluation rubric, scratchpad, exemplar.
Core patterns
| Item | Description | Example |
|---|---|---|
Role + objective | First line of system prompt names the role; second line names the win condition. | You are a senior security engineer. Goal: spot every credential leak in the diff. |
Structured output | Force JSON or labelled fields. Use response_format or json_schema. | {"verdict":"...", "issues":[...]} |
Step-back | Ask the model to articulate the question category before answering. | Before answering, classify the question as factual / opinion / advice. |
Output gating | Tell the model to answer 'INSUFFICIENT' when context is too thin instead of guessing. | If the document doesn't contain the answer, return INSUFFICIENT. |
Steering patterns
| Item | Description | Example |
|---|---|---|
Persona conditioning | Define audience, not just role. 'For a junior dev' produces different output than 'for a CTO'. | Write for a junior dev who has never used Docker. |
Anti-template | Tell the model what NOT to do. Beats vague positive instructions. | Do not write 'In conclusion'. Do not use the word 'leverage'. |
Refusal calibration | Allow the model to refuse with a structured reason. | If the request is unsafe, reply {"refuse":true, "reason":"..."}. |
Evaluation patterns
| Item | Description | Example |
|---|---|---|
Rubric in prompt | Include the scoring rubric in the prompt for evals. | Score 1-5 on: accuracy, completeness, tone. Return JSON. |
Scratchpad | Reserve <scratch>...</scratch> for chain of thought; instruct the model to put only the final answer in <answer>. | Use <scratch> for reasoning. Final reply in <answer>. |
Exemplar pinning | Pin a single high-quality exemplar with rationale at the top of the user turn. | Example of a great output: ... Why it's great: it follows the rubric exactly. |
FAQ
Do I need few-shot examples for GPT-4o?
Often no for common tasks. For strict format or domain-specific tone, yes — 2-5 well-chosen examples beat any prose instruction.
Should I tell GPT-4o to 'think step by step'?
For GPT-4o and GPT-5 non-reasoning, yes — it measurably improves multi-step tasks. For o-series reasoning models, no — they think internally.
Is JSON mode reliable in 2026?
Yes when paired with json_schema. Without a schema, JSON mode guarantees syntax but not field semantics.
Last updated: 2026-06-01.