Openai Docs Managed

Managed OpenAI Docs

Goal

Answer current OpenAI documentation questions from opened official source pages while keeping routine retrieval deterministic and model-free.

Context

Inputs To Capture

• The user's documentation question.
• The exact requested product and model names, when present.
• One optional primary route: api, codex, chatgpt, model-selection, model-migration, or troubleshooting.
• Whether the caller explicitly requests independent synthesis.

Infer only the route when it is omitted. Preserve every explicit product or model name exactly as supplied.

Constraints

Semantic Policy

• Interpret the request, select at most one primary route, and pass factual retrieval to the bundled helper.
• From this installed skill directory, create the helper's closed JSON request and run python scripts/openai_docs_retrieval.py --request <request.json>.
• The request contains exactly schema, query, requested_product, requested_model, route, and independent_synthesis; use null for an omitted product, model, or route.
• Treat an explicit model name as opaque text. Never normalize, upgrade, alias, or substitute it.
• Consume the returned record as the factual source of truth. Do not search, open, or fetch documentation again for facts already established or rejected by the helper.
• Answer only from the record's opened-page claim evidence and cite its source URLs. Preserve unresolved ambiguity and exact blockers instead of completing an unsupported answer from memory.
• Routine retrieval uses no child model. A child is eligible only when the caller explicitly requested independent synthesis or the record reports semantic_reconciliation_required; when eligible, use exactly one gpt-5.6-luna child at low reasoning effort. It may only summarize or reconcile the retained record.

Boundaries

• Use this skill for current official OpenAI product, API, model, ChatGPT, and Codex documentation, including setup, architecture, model selection, migration, and troubleshooting questions.
• Do not use it for generic software questions that merely mention an OpenAI product without needing OpenAI documentation.
• Read-only documentation retrieval requires no API key.
• The helper is the sole retrieval boundary and allows only developers.openai.com, platform.openai.com, and learn.chatgpt.com.
• Search indexes and snippets are routing evidence only. A claim is usable only after the helper opens its actual source page.
• Rank and validate required concepts through code-owned named-surface and semantic-alias sets. On one opened page, require the named surface plus a payload concept; treat unregistered phrases compositionally.
• For unspecified current-model migration, exclude older version guides from filler slots. Compare support polarity only within anchor-bearing sentences or rows.
• Within the selected route, continue past non-policy page-open failures within the coded fetch-attempt bound until the opened-page limit is reached; never select another route.
• Unavailable, disallowed-redirect, conflicting, and insufficient documentation states are blockers or ambiguity, not permission to invent an answer.

Workflow

1. Capture the exact question, product, model, optional route, and independent synthesis request.
2. Run the bundled helper once and parse its single bounded JSON record.
3. Stop factual retrieval. Use only opened-page claim evidence from the record.
4. Answer concisely with source links and disclose unresolved ambiguity or the exact blocker.

Done When

Completion Gate

• One valid helper record was consumed.
• Every factual claim in the answer maps to opened-page evidence in that record.
• Any child call satisfied the record's eligibility, model, effort, call-count, tool, and mutation constraints.
• Ambiguity and blockers remain visible without unsupported completion.

Output Contract

Return only:

• the supported answer with direct official source links
• unresolved ambiguity that affects the answer
• the exact blocker when the documentation record is not answerable
• whether an eligible independent synthesis was unavailable