Memoria jerárquica durable en CLAUDE.md por nivel usuario/equipo/módulo, con frontera explícita frente a la compaction transitoria del transcript.

Code review headless con claude -p output-format json validado contra schema; el humano sigue siendo gate final de merge.

Few-shot para calibrar bordes subjetivos con 2-4 ejemplos del mismo schema; complementa, no reemplaza, al schema.

Criterios categoricos con ejemplos positivos y negativos por severidad para reducir falsos positivos; disable temporal por categoria.

Use when controlling an agent loop by typed state and budget, including multiple tool_use coordination and Anthropic SDK Tool Runner behavior, never model prose.

Extraccion defensiva con JSON Schema, tool_choice forzado, enums con valvula de escape y nullable explicito; nunca prosa.

Slash commands vs skills: context fork, allowed-tools whitelist y argument-hint; convenciones permanentes van en CLAUDE.md.

Evaluacion critica: cross-check numerico declarado vs calculado, mismatch flag con ambos valores, sin corregir en silencio.

Mitigación material de context dilution: selección de contexto para editing/compaction y validación de ventana, thinking, media y overflow sin umbrales porcentuales inventados.

Confidence calibration contra labeled set y stratified sampling por document_type; accuracy desglosada, nunca agregada.

Seleccion de built-in tools con estrategia Grep then Read then Edit, failure modes y sin Read masivo upfront.

Investigacion adaptativa: mapeo barato, budget de exploracion acotado y re-plan disciplinado solo al invalidar la hipotesis.

System integration patterns — point-to-point, ESB, iPaaS, event mesh, API contract management, data mapping. Use when the user asks to 'design integrations', 'map system connections', 'define API contracts', 'plan event-driven integration', or mentions ESB, iPaaS, MuleSoft, API gateway, event mesh, data mapping. [EXPLICIT]

Pre-processing layer that analyzes raw user input — detecting surface errors, performing root-cause analysis (5 Whys), impact tracing (7 So-Whats), and intent gap analysis — then reformulates into a precise, actionable prompt.

Infrastructure and platform architecture — compute, network, storage, HA/DR, IAM, cloud landing zones, and cost optimization. Use when the user asks to 'design cloud infrastructure', 'plan network topology', 'define HA/DR strategy', 'set up cloud landing zones', 'optimize cloud costs', or mentions VPC, Kubernetes, serverless, multi-AZ, IAM, reserved instances, or chaos testing. [EXPLICIT]

Deploy and operate on Hostinger: shared and VPS hosting, Git/SFTP/PM2 deployment, MySQL/MariaDB, auto-renewing SSL, DNS and email via hPanel. [EXPLICIT] One of the three target platforms (Firebase + Vercel + Hostinger). Trigger: 'hostinger', 'shared hosting', 'hpanel', 'vps deploy', 'pm2', 'mariadb'.

Design current Claude Code hooks with nested matcher-entry schema, five handler types, matcher/if boundaries, permissions-aware policy, portable scripts, MCP timing checks, and deterministic validation.

This skill should be used when the user asks to create a skill from scratch, optimize an existing skill, run skill evals, or benchmark skill performance — routing to the skill-creator plugin under toolkit governance.

This skill should be used when the user asks to review a PR or diff for code quality, comments, tests, error handling, type design, or simplification — routing to pr-review-toolkit's specialized agents under toolkit governance.

This skill should be used when the user asks to scaffold a Claude Code plugin, design plugin structure, develop a command/agent/hook/MCP/skill, validate a plugin, or review a skill — routing to the plugin-dev plugin under toolkit governance.