Oracle Reference Index
Every reference/ file oracle owns, and the condition that makes it worth
reading. oracle/SKILL.md keeps only the shared-contract rows and a pointer here.
Read this when you need a reference and the Recipe registry did not already name it, or when scanning what this skill can consult at all.
| File | Read this when |
|---|---|
reference/prompt-engineering.md |
Designing prompts, deciding what a prompt cannot guarantee, triaging bad output by layer, structured outputs, Claude-specific behavior, or prompt tests. |
reference/rag-design-anti-patterns.md |
Retrieval architecture, chunking, Hybrid Search defaults, or RAG anti-pattern checks. |
reference/llm-application-patterns.md |
Choosing agent patterns, MCP design, tool-use contracts, or caching strategy. |
reference/ai-safety-guardrails.md |
OWASP LLM coverage, guardrail layers, hallucination controls, or PII handling. |
reference/evaluation-observability.md |
Building eval suites, CI gates, tracing, monitoring, or rollout checks. |
reference/cost-optimization.md |
Model routing, caching, batching, effort tuning, cost monitoring, or quality-floor gating. |
reference/llm-production-anti-patterns.md |
Production failure modes, architecture anti-patterns, MCP pitfalls, reasoning compensations. |
reference/agent-design.md |
Application-level LLM agents — tool-use loops, schemas, context/memory, delegation, termination, failure modes. |
reference/embedding-strategy.md |
RAG embedding pipeline — chunking, model selection, vector index, re-ranking, hybrid retrieval. |
reference/advanced-tool-use.md |
The tool catalog is the bottleneck (>=10 tools or >10k tokens of definitions, falling selection accuracy, aggregated MCP servers). Covers tool search, defer_loading, programmatic calling, the advisor tool. |
reference/human-ai-trust.md |
A human decides whether to accept the output — explanation design, over/underreliance and sycophancy metrics, the disagreement policy, verification affordances by risk class, escalation-to-human. |
reference/agent-behavior-contract.md |
How much role, policy, memory, relationship, and character an agent needs — necessity axes, budgets, invariant/parameter/state, spec vs prompt vs enforcement, drift. |
_common/OPUS_5_AUTHORING.md |
Sizing the design, thinking depth at DESIGN, front-loading use case/budget/tier at PROFILE. Critical: P3, P5. |