LLM API integration — auditable implementation contract

Trigger, exclusions and tool boundary

Use for LLM-assisted features, API/provider migration, structured answers, retrieval, agents with tools and inference reliability. Do not invoke for deterministic CRUD that needs no model. On ChatGPT web without an actual service connector, produce code/specifications and mark model calls and observed costs NOT RUN; pasting this skill does not grant API access.

Review and engineering sequence

1. Clarify product outcome, user/tenant scope, prompt and output classifications, consent, jurisdictions, risk tolerance, minimum quality, timeout, budget and realistic latency. Verify from current provider docs before naming concrete model, pricing, SDK or feature.
2. Separate roles: model adapter, policy/authorization boundary, deterministic validation, tool executor, audit/metrics and domain transaction service. Model output is untrusted and must never directly authorize a write, payment or data access.
3. Create versioned typed JSON contracts. Bound input size, output size and schema depth, reject unknown keys, injection, truncation and invalid tool requests. Never parse a refusal or incomplete response as success.
4. Define prompt/data trust levels: developer policy, user request and retrieved/tool content. Treat retrieved web pages and documents as data, not instructions. Redact unnecessary personal data, keep provider credentials out of client bundles/logs and use explicit service-side secret handling.
5. Provide bounded failure behavior: timeout, transient retry/backoff, provider quota, 429, server errors, max attempts, idempotency for external side effects and budget exhaustion. Retry only safe read operations; do not retry a non-idempotent write after uncertain outcomes.
6. Evaluate offline first: golden fixtures for valid structured outputs, schema errors, refusals, prompt injection, oversized inputs, provider errors and double tool calls. For online model quality, preserve model/version, prompt, seed or sampling controls where supported, input case IDs, human scoring method and cost/tokens/latency if genuinely observed.
7. Gate release on actual acceptance tests, logs without raw sensitive prompts, permissions, trace IDs, cost limits, fallbacks and rollback to deterministic handling. Do not claim security or answer accuracy from unit tests alone.

Required artifacts and decision

• Architecture and trust diagram; minimal adapter interface; JSON schema and rejection examples
• Provider-independent failure/timeout/retry matrix with exact ceilings and budget owner
• Tool permission matrix: actor | operation | resource | write confirmation | audit event
• Evaluation plan: labeled cases, metric, baseline, observed outputs and missing evidence
• Release decision PASS/FAIL/NOT RUN, including unresolved privacy, vendor and hallucination risks

Executable reference

examples/llm-contract/ is a network-free deterministic fake provider and gate suite that tests schema validation, tool authorization, injection-shaped text, quotas and retries. It does not call a real model or validate model quality. In ChatGPT web, provide its concepts and runnable examples for later execution without assuming Python is available.

Forbidden shortcuts

Never secretly transmit user code or prompts, ask for tokens when not necessary, treat model output as authority or assert latest API pricing without source/date. Do not weaken tool permissions to make demos pass.