GPT-5.6 Sol Prompting Guide

Produce the smallest prompt stack that reliably delivers the required outcome on GPT-5.6 Sol. Start with subtraction, preserve business and safety invariants, and add model-specific instructions only for measured gaps.

Quick Start

1. Identify the artifact: new prompt, review, migration, failure diagnosis, tool description, or full prompt stack. Preserve the explicitly requested model.
2. Capture the prompt contract:
   • user-visible goal and success criteria;
   • relevant context and evidence;
   • safety, business, scope, and permission boundaries;
   • tool-routing decisions and required return fields;
   • output shape, validation, fallback, and stop rules.
3. For an existing prompt, inspect available traces or checks and record the observed failure. Without execution evidence, label the diagnosis as a hypothesis; a static prompt review can still produce a useful patch.
4. Remove redundant rules, generic exhortations, obsolete process steps, irrelevant tools, ineffective examples, and contradictions.
5. Add only the smallest instruction, schema, example, or tool rule that fixes a demonstrated gap. Use references/prompt-patterns.md for task-specific patterns and assets/prompt.template.md for a new complex prompt.
6. Keep API and harness controls outside the prose prompt. Configure model, reasoning effort and mode, persisted reasoning, prompt caching, and tool eligibility in the request or runtime.
7. Run applicable existing checks. Create eval cases or launch comparison runs only when requested; compare task success, required evidence, latency, tokens, and cost on the same cases when that evaluation is in scope.
8. Deliver the final prompt, relevant runtime settings, and concise rationale for behavior-changing edits. Do not expose private reasoning.