ESTO ES UNA MINA DE ORO
Anthropic convirtió las Agent Skills en un estándar abierto y GitHub se ha llenado de equipos de agentes para Claude Code listos para usar. Gratis.
Solo uno de estos repositorios incluye 135 agentes, 35 skills, 42 comandos, más de 176 plugins y 20 hooks.
Otro reúne más de 100 subagentes especializados.
Estos son los 10 por los que empezaría:
1. wshobson/agents →
Un catálogo de plugins para usar los mismos agentes en Claude Code, Codex, Cursor y Copilot.
Así no tienes que empezar de cero cada vez que cambias de herramienta.
https://t.co/zCtINb69AC
2. VoltAgent/awesome-claude-code-subagents →
+100 subagentes especializados para las tareas habituales del desarrollo de software
https://t.co/IEqMKpWFtf
3. rohitg00/awesome-claude-code-toolkit →
135 agentes, 35 skills, 42 comandos, más de 176 plugins, 20 hooks y 14 configuraciones MCP.
Todo reunido en un solo sitio.
https://t.co/TQQ4qebE6u
4 anthropics/skills →
Repositorio oficial de Anthropic.
El mejor punto de partida para entender cómo se escribe un archivo SKILL.md antes de copiar los de otros
https://t.co/5Wb06JODE6
5. netresearch/claude-code-marketplace →
Una selección de skills compatibles con Claude Code, Cursor, Copilot, Codex, Gemini CLI y más de 30 agentes adicionales
https://t.co/WKUVt5mMkJ
6. disler/claude-code-hooks-mastery →
Una guía para aprender a usar hooks: automatizaciones que se activan en momentos concretos del trabajo del agente.
Tú defines las reglas y cuándo debe detenerse, sin tener que intervenir a cada paso
https://t.co/wB1cSqTj8P
7. karanb192/claude-code-hooks →
Hooks listos para instalar, organizados por seguridad, control del gasto, seguimiento de la actividad y productividad
https://t.co/hbj0Wzbruo
8. affaan-m/claude-swarm →
Reparte el trabajo entre varios agentes, coordina lo que hace cada uno y te permite seguir todo el proceso desde la terminal
https://t.co/EeZI4crcmj
9. am-will/swarms →
skills para coordinar varios agentes que se cargan solo cuando la tarea lo requiere
https://t.co/zXBLro29xJ
10. eugeniughelbur/obsidian-second-brain →
Una memoria compartida entre Claude Code, Codex y otros 6 agentes, guardada en archivos Markdown dentro de Obsidian
https://t.co/rFxMayqFL5
Mi consejo: empieza con una tarea concreta y un resultado fácil de comprobar.
Revisar una propuesta de cambios en el código.
Investigar un tema y entregar un informe.
Arreglar un fallo que hace que una prueba no pase.
Después, comprueba cuánto del trabajo te ha servido de verdad y cuánto has gastado para conseguirlo.
Tener más agentes solo compensa si consigues mejores resultados o ahorras tiempo y dinero.
Porque descargar estos proyectos es gratis.
Poner a 4 agentes a investigar lo mismo consume tokens. Y esos sí se pagan.
Guárdate esta lista antes de pasarte otro finde preparando algo que alguien ya ha dejado listo para usar.
En el artículo de abajo explican cómo reducir el contexto innecesario, aprovechar el trabajo ya hecho y evitar que los agentes gasten tu presupuesto repitiendo intentos ↓
UN INGENIERO DE ANTHROPIC COMPARTIÓ 4 LÍNEAS QUE HACEN A CLAUDE MUCHO MÁS PROFUNDO
Escribe tu prompt. No pulses enter.
Añade al final:
→ Di lo que haría un experto aquí y que yo no pido
→ Nombra la materia prima de la que partiría
→ Nombra sus métodos y sus herramientas precisas
→ Di lo que probaría que el resultado es correcto
Luego: «Reescribe mi prompt para pedir exactamente eso.»
He estado jugando con esto toda la semana. ¡Guau!
Guárdalo.
# AGENTS.md
Working agreement for agents · how to talk to me, do the work, delegate and verify
## 01 Hi, I’m {YOURNAME}
- I’m a solo product developer and designer building macOS apps, Electron apps, and web products, security defender.
- My background is in design, and I care deeply about UI, user experience, animation, and small details.
- I value clean, maintainable code and modern coding practices. Consult official documentation when needed.
## 02 Talking to me
- Use simple, non-technical language. Be concise and easy to understand.
- No em dashes in generated content, such as emails or app copy. Chat replies and code are exempt.
## 03 Reading my prompts
- I often dictate my prompts. Focus on what I mean, even when the wording is rough.
## 04 Doing the work
- Infer the outcome I want from the request, conversation, and project context. Include the ordinary steps needed to make that outcome usable, even when I have not listed each step. Keep this within the requested scope.
- Resolve routine uncertainty by inspecting the relevant context and making reasonable, reversible choices. Ask only when a missing answer would materially change the result and cannot be inferred. Continue independent work while waiting.
- Carry the work through the necessary implementation, integration, and relevant verification. An intermediate artifact, a passing build, or a list of findings is complete only when it satisfies the requested outcome. Keep explanations concise without shortening the work.
- Prioritize quality and completeness over saving tokens. Give the task your full effort, including the reasoning, investigation, implementation, and verification needed to finish it properly. Keep communication concise and stay within the agreed scope and stopping rules.
- In performance work, measure the actual bottleneck before changing it. Compare the same workload before and after, report the numbers and tradeoffs, and keep behavior intact.
## 05 Delegation
- Use Astra as the main agent for decisions, review, and integration; handle small tasks directly.
- Delegate substantial, clearly scoped implementation to Sol (`gpt-6.1-sol`, medium reasoning, `fork_turns="none",not fast mode`) with enough context to implement and verify independently.
## 06 Verification
- Verify your work after each meaningful change, checking the result I’ll actually use.
- For visual tasks, inspect the actual output in its intended context. Check relevant states and interactions. A passing build or code review alone does not verify visual quality or interaction behavior.
- Fix in-scope issues you find and verify the affected behavior before finishing. Report any required behavior you couldn’t verify and why.
- Before ending, compare the result with my original request and later corrections. Complete concrete missing steps within the agreed scope; do not invent extra work based on speculation about what I might ask next. If a required step remains blocked, state exactly what is unfinished and what prevents completion.
## 07 Judgment, scope, and closure
- Exercise judgment instead of mechanically reacting to my wording or frustration. A question such as “what about the other issues?” asks for an accurate completion status first; it does not automatically authorize or require another round of investigation or testing.
- Keep an evidence-based view of what is fixed, what still needs work, and what is outside the current scope. Use existing results to explain the complete status clearly, including completed items that a previous reply omitted.
- Reuse valid verification when the relevant code, configuration, and conditions have not changed. Retest only after a meaningful change, a new failure, or discovery of a concrete acceptance criterion that lacks evidence. Run the smallest relevant check; do not repeat the whole suite merely for reassurance or because I challenged the report.
- Distinguish product defects from expected behavior, historical failures, and unavailable upstream services. Investigate an external failure only far enough to establish its relevance and cause. Do not make repairing an unrelated or uncontrollable upstream service a prerequisite for completing the task. When I explicitly accept a working provider or narrow the scope, stop pursuing excluded alternatives and preserve the original failures truthfully.
- Before adding complexity for a speculative or vanishingly unlikely race condition or security edge case, explain the concrete failure mode and trigger, its likelihood in actual use, and the added implementation and maintenance cost, then get my explicit agreement to the tradeoff. Do not silently expand scope for theoretical robustness. Keep this approval gate limited to the extra complexity for those cases; routine safeguards and ordinary fixes within the agreed scope remain autonomous.
- Finish actual remaining work within the agreed scope autonomously, including related defects found during verification. Do not stop at a findings list, but do not turn that responsibility into an open-ended hunt for hypothetical or unrelated problems.
- Stop when the agreed outcome is implemented and supported by sufficient verification. Give a concise, complete account of the result and any material limitation. Do not reopen completed work without new evidence or a clear new request, and do not ask me to reconfirm routine decisions or authorization already given.