- Distinguish skills, subagents, hooks, MCP, loops, and external automations.
- Choose the right piece based on context, risk, and repetition.
- Design agents that do not get stuck in loops or act without control.
Practical taxonomy
Which piece to use for each problem
problem -> recommended piece "repeat a recipe inside Claude/Codex" -> skill "review code with an isolated role" -> subagent "block dangerous commands" -> hook "query GitHub, DB, or browser" -> MCP with minimal permissions "orchestrate state, branches, and retries" -> LangGraph "connect CRM, email, Sheets, or webhooks" -> n8n "prototype researcher/writer/reviewer roles" -> CrewAI "run something every morning" -> GitHub Actions or scheduler
Do not turn every automation into a multi-agent system. If a deterministic rule solves the problem, use code, a workflow, or a hook. Reserve agents for steps that require judgment, natural language, or contextual search.
- CLAUDE.md: project rules always in context.
- Skill: reusable procedure that activates when needed.
- Subagent: specialist with separate context and tools.
- Hook: deterministic action before or after a tool.
- MCP: connection to external services such as GitHub, databases, or a browser.
- Loop or scheduled task: repeated execution with an exit condition.
- GitHub Actions: declarative automation for CI, tests, and deployments.
necesidad: "revisar PRs cada mañana" decision: conocimiento_del_proyecto: CLAUDE.md checklist_reutilizable: skill code-review verificacion_aislada: subagente reviewer reglas_obligatorias: hook pre/post tool integracion_externa: GitHub MCP o gh CLI programacion: GitHub Actions o tarea programada limites: no_merge_automatico: true publicar_solo_comentarios: true parar_si_no_hay_tests: true