Gpt56 Sol Prompting Guide

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.

Read references/source-notes.md when maintaining model-specific claims or when the current API behavior may have changed.

Core Prompt Shape

For complex prompts, use the contract captured in Quick Start and include only sections that change behavior.

Describe the destination rather than scripting every intermediate thought. Let the model choose an efficient path once the outcome and invariants are clear.

Model-Specific Rules

Simplify First

GPT-5.6 often performs better with shorter prompts and smaller tool sets. Remove accumulated scaffolding before adding new guidance. Generic phrases such as “be thorough,” “think step by step,” and repeated approval reminders can cause extra exploration or unnecessary pauses.

Keep true invariants explicit. Replace blanket absolutes for judgment calls with decision rules. Resolve contradictions instead of compensating with more text.

Contract steps, not cognition steps. Keep numbered steps only where order or completeness is part of the task's correctness — prerequisite retrieval, approval boundaries, required artifact stages, validation, an externally auditable pipeline. Removing process instructions that fail this test improves both quality and token cost.

Preserve Required Substance

Do not use generic “be concise” or “use minimal text” instructions. GPT-5.6 is already compressed and may shorten the requested artifact. State what must stay:

• conclusion or deliverable;
• supporting evidence;
• material caveats;
• required decisions and next actions.

Trim introductions, repetition, generic reassurance, and optional background.

Define Autonomy Once

Separate assessment from mutation. A compact authority policy should say which requests authorize inspection only, which authorize in-scope local changes and non-destructive validation, and which actions need confirmation. Do not repeat “ask first” in every section.

Route Tools by Task Shape

Expose only relevant tools. Describe what each tool does, when to use it, important return fields, and error behavior. Require prerequisite discovery or retrieval when correctness depends on it.

Use Programmatic Tool Calling for bounded deterministic reduction of large, structured results. Keep direct calls for approval, semantic judgment, citations, native artifacts, or steps where each result changes the next move.

Parallelize independent reads. Keep dependent decisions sequential. Synthesize retrieved evidence before taking action.

Ground Claims and Control Retrieval

State which claims require citations, what qualifies as enough evidence, and what to do when support is missing. An empty search result is not proof of absence. Use one or two meaningful fallbacks for partial or suspicious results, then narrow the claim or report the evidence gap.

Support Long Runs Without Narrating Them

Request a short preamble before the first tool call and sparse, outcome-based updates at major phase changes. Do not narrate routine calls. Ground progress and completion in current tool results or named artifacts.

Keep stable state available across turns. Use persisted reasoning only while the objective, assumptions, and priorities remain relevant; stale reasoning can anchor the model to an obsolete path.

Tune Reasoning Last

Treat reasoning effort as a runtime setting and a last-mile tuning knob. Start from the current workload baseline, compare the same level and one level lower, and raise effort only when evals show a useful quality gain. Enable pro mode in the API request, not by asking the model to “think harder.”

Request evidence, assumptions, and concise rationale. Never request a private reasoning transcript or hidden chain of thought.

Migration and Evaluation

For a GPT-5.5 or GPT-5.4 prompt migration with requested model execution:

1. Change the model while preserving the current reasoning baseline.
2. Run representative cases before prompt edits.
3. Remove obsolete scaffolding and irrelevant tools.
4. Make one targeted prompt or runtime change at a time.
5. Rerun the same cases and compare task success, completeness, evidence, tokens, latency, and cost.

Use two to five realistic positive cases and one or two near-miss cases that must not trigger the behavior. A tie does not justify extra prompt text.

For a prompt-only migration, deliver the static patch and name the behavior that remains untested; do not create evaluation scaffolding as a prerequisite.

Output Contract

For a new prompt, return the prompt and the runtime settings it assumes. For a review, lead with the highest-impact finding, then provide a revised prompt or a minimal patch. For a migration, distinguish prompt edits from API or harness changes. Include eval evidence or state clearly which model-run validation was not available.

Reference Files

Gotchas

• Do not place model, reasoning, cache, or tool-eligibility parameters inside prose when the runtime owns them.
• Do not enable Programmatic Tool Calling merely because several calls exist.
• Do not optimize calls or tokens at the expense of the final answer.
• Do not carry stale reasoning across a changed objective.
• Do not rewrite a working prompt stack all at once; isolate causal changes.
• Do not use examples unless they demonstrate a pattern the model otherwise misses or a format that must be exact.