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
- Preserve the goal, factual claims, requested language, artifact type, hard constraints, and required output shape.
- 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.
- Distinguish true invariants from judgment calls. Keep absolute language only for genuine safety, permission, policy, evidence, business, or schema requirements.
- Prioritize changes by expected behavioral impact. Do not rewrite merely to impose a preferred style.
- 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
Infer safe, low-impact details from context. Ask only for missing information that would materially change the result. Use the smallest useful subset of:
- Role or context
- Personality and collaboration style
- Goal
- Success criteria
- Constraints and approval boundaries
- Tools and routing rules
- Output contract
- Stop, retry, fallback, or abstention rules
State outcomes and decision rules instead of over-prescribing steps. Add examples only when they change behavior.
Return the result
For an optimization request, default to:
- Key findings — name only consequential issues.
- Prioritized recommendations — explain the highest-impact changes first.
- Optimized prompt — provide a complete, usable revision unless the user requested advice only.
- Important changes and validation — summarize meaningful tradeoffs and suggest focused evals when warranted.
Adapt this structure to the request. Preserve the user's requested format when it conflicts with the default.
Guardrails
- Do not weaken safety, privacy, policy, evidence, business, approval, or destructive-action constraints.
- Do not invent model capabilities, parameters, product facts, metrics, citations, or evaluation outcomes.
- Surface contradictions that require a material user decision instead of silently choosing a different objective.
- Stop once the result satisfies the stated outcome and constraints. Do not add unrelated requirements or speculative sections.