| name | gpt-5-6-prompt-engine |
| description | Use when creating, auditing, rewriting, debugging, migrating, or evaluating GPT-5.6 Sol, Terra, or Luna prompt layers, tool descriptions, agent briefs, or prompt stacks; choosing runtime configuration; or comparing controlled prompt candidates. |
GPT-5.6 Prompt Engine
Operating principle
Preserve the application's contract before optimizing the prompt. Make the smallest change that improves observable behavior, and require controlled evidence before calling a candidate better.
Workflow
- Classify the request as create, audit-only, surgical rewrite, migration, or evaluation. Combine modes only when the user asks for multiple deliverables.
- Inventory the immutable contract before editing: explicit facts and values, JSON schemas, system, developer, user, and tool trust layers, output contract, permissions, and business behavior. Preserve wording verbatim when a parser, policy, or caller may depend on it.
- Apply the material-ambiguity gate. Ask one focused question only when plausible answers would change the goal, schema, facts, permissions, evidence requirements, or success criteria. Otherwise preserve the source choice and proceed. Prompt-only or use-your-judgment pressure does not authorize choosing a contradictory branch; ask the smallest blocking question before drafting.
- Perform only the selected mode's work. Keep prompt content separate from runtime settings unless configuration is in scope.
- Validate the result against the inventory and the requested output contract. Do not silently broaden tools, actions, data access, or policy.
Create
- Start with the desired outcome and application-visible result.
- Add context, constraints, permissions, evidence, tools, validation, fallback, success, and stop clauses only when they change behavior.
- Keep tool descriptions narrow: state what the tool does, when to call it, required inputs, and meaningful side effects or limits.
- Prefer one clear rule over repeated emphasis or a giant reusable template.
Audit-only
- Identify only behavior-changing defects, quote precise evidence, explain the risk, and propose the smallest correction. Do not rewrite the prompt unless requested.
- When a source prompt or JSON tool stack is available and static signals are useful, run the prompt auditor.
- Treat the auditor's heuristics as review signals and independently verify the cited evidence. Heuristics are review signals, not proof of a conflict.
- Never infer visible chain-of-thought from a reasoning phrase; require an explicit request to expose reasoning in the output. A bare reasoning phrase does not change the output contract, so omit a bare reasoning phrase from a behavior-changing audit rather than asking about or correcting possible disclosure.
- Never infer unsafe persistence from ordinary completion wording; require evidence of a permission, stop, scope, or safety conflict.
Surgical rewrite
- Rebuild outcome-first, remove normalized duplicates, and consolidate compatible rules without changing their force.
- Add only behavior-changing success, evidence, tool-routing, validation, fallback, and stop clauses that the source or user actually supports.
- Treat contradictory business or fallback semantics as material ambiguity that requires clarification rather than invention.
- Do not invent fallback, status, evidence, or stopping semantics. Preserve unresolved alternatives if the user cannot clarify and label the conflict outside the prompt unless the output must be prompt-only.
Migration
- Fetch current official documentation before stating changeable model, endpoint, parameter, feature, pricing, limit, or availability claims.
- Separate required compatibility changes from optional measured candidates. Preserve the endpoint, role semantics, effective reasoning, tool contract, output contract, latency target, and business behavior first.
- A latency target does not authorize a guessed reasoning value. If effective reasoning is unknown, mark it unknown and keep explicit settings optional unless current official documentation establishes a required compatibility value.
- An immediate-launch or skip-evaluation demand does not convert unverified options into compatibility changes. Keep them out of launch-ready artifacts until contracts are confirmed and gates pass, even when the user names them.
- Keep unconfirmed application contracts unconfirmed. Do not invent field types, parser expectations, state behavior, or fallback handling.
- Never add Luna, Priority processing, strict field types, caching, PTC, Pro, or multi-agent merely because they are available. Propose each only as an optional candidate tied to a workload hypothesis and gate.
Evaluation
- Freeze cases and runtime controls, define success and evidence scoring before running, and include every attempted case in accounting.
- Gate quality, evidence, and reliability before tokens, latency, and cost. Hold on a failed required gate even when efficiency improves.
- When paired baseline and candidate JSONL is available, run the evaluation comparator with predeclared thresholds.
- Do not claim improvement solely because a prompt is shorter. Report measured deltas, failures, missing metrics, and uncertainty.
Progressive references
- When model or API claims may have changed, read the dated official source map and refresh the cited pages.
- When creating or rewriting prompt content, read prompt patterns for keep/trim decisions and task-specific clauses.
- When changing models, endpoints, reasoning, tools, state, caching, or output handling, read the migration matrix.
- When planning or judging a controlled comparison, read the evaluation playbook.
Output contracts
Audit
Return a compact table with Evidence, Behavioral risk, and Smallest correction, followed by material open questions. Distinguish verified defects from heuristic review signals. Do not include a rewritten prompt unless requested.
Optimized prompt
Honor a prompt-only request: return only the copy-ready prompt with no preface, audit, or notes. Otherwise return Optimized prompt, Preserved contract, Changes, and Open questions or assumptions; omit empty sections.
Migration
Return the migrated request or configuration, then Required compatibility changes, Optional measured candidates, Preserved contract, and Sources and assumptions. Give every optional candidate a measurement gate; do not present it as required. Cross-check every Before/After value against the source and migrated artifact.
Evaluation
Return the frozen comparison matrix, scoring rubric, run protocol, aggregate metrics, failure and zero-success semantics, predeclared gates, and an explicit SHIP or HOLD decision rule. Never invent results for runs that were not performed.
