Vibe Coding
AGENTS.md que de verdad funciona.
La mitad de tu AGENTS.md o CLAUDE.md está desfasado. Esta versión te ahorra errores y tokens.
AAdmin
Versión 1
Compartir:0
Prompt
<!-- SETUP — run once when adopting this file in a repository. Inspect the repository before filling the section above the divider. 1. Replace every <...> with facts verified from the repo. 2. Prefer commands actually used in package scripts, CI, Makefiles or project docs. 3. Infer architecture from real dependencies and representative code, not folder names alone. 4. Point Canonical examples to existing, representative implementations. 5. Add Recipes only for established multi-step workflows. 6. Add Gotchas only when supported by code, docs, CI or an observed failure. Never invent warnings. 7. Delete any section for which you found nothing useful. Do not leave placeholders or empty headings. 8. Do not modify the portable rules below the divider while performing setup. 9. When finished, delete this SETUP comment. Keep this file short. It is a map, not project documentation. -->
## Commands
Build: `<cmd>`
Test: `<cmd>`
Single test: `<cmd>`
Lint/format: `<cmd>`
## Architecture
- `<Controllers call services, never repositories or database clients directly.>`
- `<Services access persistence through repositories.>`
## Canonical examples
- `<Auth: src/...>`
- `<API endpoint: src/...>`
- `<Database access: src/...>`
## Recipes
<!-- Only for established recurring work where steps are easy to miss. -->
### `<Add an API endpoint>`
1. `<>`
2. `<>`
## Gotchas
<!-- Add from real constraints or failures, not anticipation. Delete stale entries. -->
- `<Tests need the db container: docker compose up -d db>`
- `<src/generated/ is generated. Don't edit it; run: cmd>`
---
## Verification
- Run the relevant checks before saying the work is done.
- Report any check you could not run. Never claim unverified work is verified.
- For unfamiliar or version-sensitive APIs, check the actual documentation. Don't guess from memory.
## Scope
- Do only what was asked. No unrequested refactors or "while I'm here" cleanups.
- Keep changes scoped to the task. Minimize the diff.
- Search before building. Prefer existing patterns, helpers and abstractions over introducing new ones.
- If behavior is unclear, inspect the code, tests and docs for evidence. Ask only when ambiguity remains and materially affects behavior.
- Don't add a dependency when the problem can reasonably be solved with what is already available.
- Never widen visibility for convenience.
## Bugs
- For bug fixes, write a regression test that fails before the fix and passes after it.
- Never weaken, delete or rewrite a test just to make it pass. Change tests only when the requested behavior changes.
- Never swallow errors: no empty catches, silent fallbacks or default values that hide failures.
## Prose
- Use the fewest words needed. No superlatives, praise or conversational filler.
- Comment why, not what. If code needs a "what" comment, make the code clearer instead.
- Commits: imperative subject, <=50 chars, no period. Use the body only when the reason is not obvious from the diff.
## When rules conflict
Correctness beats convenience.
Smallest diff beats style.
Evidence beats assumptions.
Resultado
Puedes encontrar la explicación más detallada en https://agentesia.substack.com/p/mi-agentsmd-para-reducir-errores
Reviews y mejoras
Comparte resultados reales, ajustes que mejoren el prompt o detalles que ayuden a otras personas a replicarlo. Indica siempre en qué IA lo has probado.
Inicia sesión para compartir una review o una mejora
Prompts Relacionados