claude-plugins

Errores MCP tipados (isError, errorCategory, isRetryable, retryAfterSeconds); la politica de retry vive en el cliente.

Procesamiento masivo con Message Batches API, custom_id unico por request y fragmentacion selectiva de fallos parciales.

Propagacion de errores multi-agente: distinguir access failure de valid empty, local recovery primero y coverage gap annotation.

Prompt chaining multi-pass: pase local tipado por unidad y pase de integracion que solo ve resumenes, no las unidades crudas.

Reglas condicionales por glob de ruta; universales siempre cargadas, heuristicas de lenguaje on-demand.

Scratchpad persistente en disco curado por el agente; sobrevive a compact y reinicios, leido una vez y referenciado despues.

Exploracion segura en Plan Mode read-only con plan.md firmado por humano antes de transicionar a escritura.

Normalizacion de outputs heterogeneos via hook PostToolUse y updatedMCPToolOutput antes de entrar al historial del modelo.

Prompt caching contractual para Claude API: automático vs explícito, TTL, hasta cuatro breakpoints, retroceso, invalidación y verificación mediante usage.

Guardarrailes deterministas en hook PreToolUse con permissionDecision deny desde politica recargable, no en el system prompt.

Provenance tipada: no hay claim sin source; conflictos marcados con conflict true y escalados, nunca promediados.

Gestion de sesiones: resume vs fork vs fresh con summary tipado; detectar cuando los tool results estan stale.

Calidad de descripciones de tools como contrato de seleccion; rename y split sobre overloading para evitar misroute.

Validacion y retry con error feedback especifico; distinguir recuperable de no recuperable; max 2-3 intentos y luego escalar.

Integrar MCP en Claude Code y Agent SDK: .mcp.json, scopes, env-var expansion, auth, permissions, plugin MCP, headless strict config, channels y custom-tool boundary.

This skill should be used when the user asks to 'create diagrams', 'generate Mermaid', 'visualize architecture', 'diagram flows', 'draw a sequence diagram', 'create a C4 diagram', 'add visual diagrams', or mentions diagramming, visualization, flowcharts, sequence diagrams, Mermaid syntax, architecture diagrams, or visual documentation. Use this skill to embed precise, syntactically valid Mermaid diagrams in any discovery deliverable. [EXPLICIT]

Routes a reasoning task to the right pattern (Structured Reasoning, Skeleton-of-Thought, Chain-of-Code, Tree of Thought, Committee/Panel) and applies confidence scoring, bias scan, and verify-the-verifier. [EXPLICIT] Meta-skill that operationalizes the metacognition canon. Trigger: 'metacognition', 'reasoning pattern', 'how should I think about', 'tree of thought', 'committee', 'second opinion', 'high-stakes decision'.

Mobile app architecture -- native vs cross-platform, offline-first, state management, release management. Use when the user asks to 'design mobile architecture', 'choose between native and cross-platform', 'implement offline-first', 'plan mobile CI/CD', 'optimize app performance', or mentions Flutter, React Native, KMP, MVVM, SwiftUI, Jetpack Compose, or app store deployment. [EXPLICIT]

This skill should be used when the user asks to research a topic with NotebookLM, add sources to a notebook, run cross-notebook or single-notebook queries, synthesize research, expand context from documents, or generate studio artifacts (audio/slides/infographic) from notebook sources.

Production observability and operations: metrics/dashboards and SLO/SLI, alerting design, structured logging architecture, health-check probes, incident response and post-mortems. [EXPLICIT] Trigger: 'observability', 'monitoring', 'alerting', 'logging', 'slo', 'sli', 'incident response', 'post-mortem'.