Optimize GPT-5.6 Prompts

Start from the user's outcome

Determine whether the request is to diagnose, optimize, migrate, or draft from scratch. Identify the intended user-visible result, success criteria, required evidence, permission boundaries, output contract, and true invariants before changing wording.

Read references/gpt-5-6-guidance.md for the detailed review criteria. Read only the sections relevant to the request. Treat it as prompting guidance, not as a source for current API limits, pricing, parameters, or feature availability; retrieve current official OpenAI documentation when those facts matter.

Review an existing prompt

1. Preserve the goal, factual claims, requested language, artifact type, hard constraints, and required output shape.
2. Find the smallest set of material problems: repetition, contradictions, obsolete scaffolding, vague outcomes, missing completion criteria, unnecessary process rules, irrelevant tools, weak stopping conditions, or unsupported claims.
3. Distinguish true invariants from judgment calls. Keep absolute language only for genuine safety, permission, policy, evidence, business, or schema requirements.
4. Prioritize changes by expected behavioral impact. Do not rewrite merely to impose a preferred style.
5. When behavior may change, recommend representative evals and isolate changes so results remain attributable.

If the prompt is already effective, say so and make only targeted changes.

Draft a prompt from scratch