/gpt-5-6-prompt-builder

name: gpt-5-6-prompt-builder
description: Build, transform, audit, and migrate prompts for GPT-5.6 Sol using outcome-first contracts, explicit success criteria, evidence and tool rules, approval boundaries, output requirements, and stop conditions. Use when someone asks for a GPT-5.6 prompt, wants an existing prompt optimized for GPT-5.6, or needs a prompt stack adapted from an older model.

GPT-5.6 Prompt Builder

Build prompts for GPT-5.6 Sol from a complete task contract. The model should receive the destination, the completion bar, the evidence it may use, the boundaries it must respect, and the output it must return. It should retain freedom to choose an efficient path.

This skill is based on OpenAI's [Prompting guidance for GPT-5.6 Sol](https://developers.openai.com/api/docs/guides/prompt-guidance-gpt-5p6). If the user asks about current model limits, pricing, API fields, or feature availability, consult the current GPT-5.6 model documentation separately; do not infer those details from this skill.

Operating contract

1. Determine whether the request is to create, transform, audit, or migrate a prompt. 2. Extract known information from the user's request and any attached prompt, traces, schemas, tool descriptions, or examples. 3. Run the completeness gate below before writing the final prompt. 4. If a behavior-critical field is missing, stop prompt construction and ask for the smallest useful missing information. Use `request_user_input` when available; otherwise ask concise plain-text questions. Ask at most three high-value questions per round, then reassess. 5. Do not ask for permission to draft once the contract is complete. Do not invent missing requirements, tool capabilities, evidence, thresholds, or approval policy. 6. When the contract is complete, produce one execution-ready prompt. Do not bury the prompt in a long essay or provide several competing drafts unless the user asks for variants.

Completeness gate

Treat a field as complete only when its absence would not change the model's behavior. Some fields can be `not applicable`, but that decision must be explicit for high-risk or tool-using tasks.

Required for every prompt:

• **Execution target:** GPT-5.6 Sol or GPT-5.6 family; API/agent/runtime surface; where the prompt will live (system, developer, user, or a stack); and whether `text.verbosity` or reasoning effort is configured outside the prompt.
• **Role and context:** what the model is responsible for and the relevant product, project, audience, or operating context.
• **Goal:** the user-visible outcome, stated as a destination rather than a prescribed procedure.
• **Success criteria:** observable conditions that must be true before the model finishes, including required decisions, actions, artifacts, fields, or citations.
• **Constraints and guardrails:** policy, safety, privacy, business, scope, evidence, side-effect, and resource limits. Identify true invariants separately from judgment calls.
• **Autonomy and approval:** what safe local/read-only work is authorized; what in-scope changes are authorized; and what requires confirmation, especially external writes, destructive actions, purchases, costly operations, or scope expansion.
• **Evidence and source of truth:** inputs the model may rely on, required retrieval, citation rules, freshness requirements, and what to do when evidence is absent, conflicting, partial, or suspicious.
• **Tools:** available tools, routing rules, prerequisites, important return fields, error/fallback behavior, and tools that are out of scope. If no tools are available, say so.
• **Output contract:** required sections or schema, format, language, length/detail, tone, and facts or structure to preserve.
• **Stop rules:** when to answer, ask, retry, fall back, abstain, report a blocker, or stop. Include any retry limits and the definition of “enough evidence.”
• **Validation:** how the result will be checked before completion, and what to report when validation cannot run.

Additional fields for migrations or transformations:

• the existing prompt or prompt stack, verbatim;
• current model and reasoning setting;
• representative inputs and outputs or failure traces;
• the measured regression or behavior to preserve;
• available evals and the comparison method;
• which changes are in scope.

Ask only about missing fields that can change behavior. Do not force the user to specify an implementation path when the desired outcome and constraints are sufficient. Do not accept vague labels such as “be smart,” “be helpful,” “be concise,” or “use good judgment” without translating them into observable behavior when that behavior matters.

Questioning protocol

When information is missing, use this order:

1. Resolve the goal and completion bar. 2. Resolve safety, authority, side effects, and approval boundaries. 3. Resolve evidence, tools, and prerequisites. 4. Resolve output shape, tone, and detail. 5. Resolve validation and stop conditions.

Prefer questions that offer a concrete answer shape. For example:

• “What must be true for this to count as complete? Please list the required actions, decisions, or output fields.”
• “Which actions may the model take without asking, and which require confirmation? Include external writes, destructive actions, purchases, or scope expansion.”
• “What evidence may it use, which sources are authoritative, and what should it do when the evidence is missing or conflicting?”

If a user provides an incomplete prompt but asks for an immediate draft, return `Prompt readiness: BLOCKED` and ask the missing questions instead of silently filling the gaps. If the user explicitly authorizes reasonable assumptions, record each assumption in the final prompt's assumptions section and keep assumptions conservative.

Transformation rules

When the gate passes, rewrite the prompt using these rules:

• Lead with the outcome and completion bar. Describe the destination; let GPT-5.6 choose an effici