AI Engineering

This skill is the router for AI changes in any codebase using it. Keep it loaded for LLM calls, agents, workflows, or structured AI schema changes, then read only the subskill files that match the task.

Quick Fetch Map

• Direct LLM calls: subskills/llm-calls.md direct-call fit lines 5-14; call defaults lines 16-25; rate limits, retries, and malformed-output repair lines 26-30; provider caching lines 35-64; prompt shape lines 66-75.
• Schema design: subskills/schema-design.md core rules lines 18-31; field ordering and routing patterns lines 32-47; minimal examples lines 48-85; tool schema rules lines 93-123; citation/confidence schema patterns lines 124-153.
• Schema references: subskills/schema-examples.md Python examples lines 16-109; TypeScript examples lines 110-196; SGR parameters lines 197-211. subskills/schema-provider-support.md portable baseline lines 5-17; provider matrix lines 19-32; portability rules lines 34-40.
• Retrieval and grounding: subskills/retrieval.md architecture choice lines 5-18; retriever rules lines 20-31; chunking lines 33-43; grounded answer design lines 45-62; retrieval evals lines 64-80.
• OpenAI web search: subskills/openai-web-search.md when to use lines 17-21; Responses API settings lines 23-33; URL annotations lines 35-46; structured-output pattern lines 48-138; claim-level citation mapping lines 140-165; source quality lines 167-173.
• Citations and provenance: subskills/citations.md two-stage architecture lines 23-41 and 63-89; page markers lines 90-169; geometry lines 170-182; schemas lines 183-224; matching heuristics lines 225-245; UX/evals lines 246-279.
• Confidence and review routing: subskills/confidence.md output shape and field order lines 41-85; confidence criteria lines 98-156; workflow, mapping, and contradictions lines 157-206; lean/audit schema tradeoff lines 207-280; calibration lines 281-300.
• Workflows: subskills/workflows.md workflow definition lines 5-14; complexity order lines 16-27; workflow patterns lines 29-53; rules, retries, and cache-aware scheduling lines 55-65; verifier pattern lines 67-76.
• Agents: subskills/agents.md when to use agents lines 27-38; agent rules and budgets lines 39-78; tool design lines 79-92; context and multi-agent cautions lines 93-110; loop examples lines 121-168.
• Tracing: subskills/tracing.md default stack and storage lines 5-30; trace boundaries lines 32-47; Python and TypeScript setup lines 49-73; design rules and trace search lines 75-94; minimal patterns lines 96-109.
• Safety and evals: subskills/safety-evals.md security rules lines 5-14; model selection lines 16-24; eval flywheel and rules lines 26-51; metrics and calibration lines 53-76; dataset and human review lines 78-98.
• Shared references: source index sources.md; trigger eval prompt set evals/trigger-queries.json.