Productividad del desarrollador

Writing for Agents

Matt PocockMITIdioma del Skill: Inglés
Licencia verificadaSin scripts
DemostraciónEjecución real · claude-opus-5-5 · octubre de 2026
Can you improve this CLAUDE.md for our repo? Agents keep ignoring parts of it. ```markdown # Project Guidelines This is a Next.js 15 app using the App Router with TypeScript. We use pnpm as our package manager. The scripts are in package.json. ## Important Rules - Please always make sure to be very careful and thorough when making changes. - Don't use npm or yarn. Don't use npm or yarn under any circumstances, always use pnpm. - Never ever use the `any` type. - Run `pnpm lint` and `pnpm test` before you finish. - Don't put business logic in React components. - Make sure you write good code that follows best practices. ## Database We use Prisma with PostgreSQL. The schema is at prisma/schema.prisma. When you change the schema you need to run `pnpm prisma migrate dev --name <name>` and you also need to run `pnpm prisma generate` afterwards, otherwise the types will be out of date and things won't compile. Don't forget to run prisma generate!! ## Testing We use Vitest. Tests live next to the code in `*.test.ts` files. Run `pnpm test`. Please don't forget to run tests before finishing. ## Style Use Tailwind for styling. Don't write CSS files. Don't use inline styles. Don't use styled-components. ```
Skill: writing-for-agents

The main problem is that the important rules are hard to spot among filler. Here's a rewrite:

# Project Guidelines

Next.js 15 (App Router) + TypeScript, managed with **pnpm**. Scripts are in `package.json`.

## Rules

- Run every package and script command through `pnpm` (`pnpm add`, `pnpm dlx`, `pnpm <script>`). The lockfile is `pnpm-lock.yaml`, and running another package manager corrupts it.
- Use precise types everywhere. For data whose shape you don't know yet, use `unknown` and narrow it with type guards or a schema. Never use `any`.
- Keep React components presentational. Put business logic in <!-- TODO: e.g. `src/lib/` or `src/server/` --> and call it from the component.
- Style with Tailwind utility classes only.

## Database (Prisma + PostgreSQL)

After any edit to `prisma/schema.prisma`, run both commands, in this order:

1. `pnpm prisma migrate dev --name <descriptive_name>`
2. `pnpm prisma generate`

If you skip step 2, the generated types go stale and the build fails in ways that are hard to trace back.

## Testing

Vitest. Put each test next to its source file as `<name>.test.ts`.

## Done means

You're finished only when both commands pass with zero errors on your final change:

- `pnpm lint`
- `pnpm test`

What I changed and why:

  1. Cut lines that don't change behavior. "Be careful and thorough" and "write good code that follows best practices" are things the agent already tries to do. They make the file longer and draw attention away from the rules that matter.
  2. Gave each rule one home. "Run tests" appeared twice and "use pnpm" was said twice in one bullet. When an instruction is repeated, it's hard to tell how important it is compared with the rest. Lint and test now live in a single Done means section at the end, where the agent checks whether it's finished.
  3. Said what to do, not just what to avoid. Rules like "Don't use npm or yarn", "Never use any", and "Don't write CSS / inline / styled-components" put the banned thing front and center. Each rule now names what to use instead (pnpm, unknown + narrowing, Tailwind classes). I kept "Never use any" as a hard limit, paired with the alternative.
  4. Added the reasons agents can't figure out on their own. The lockfile damage and the stale Prisma types are why those rules exist. A rule that comes with its reason is less likely to be dropped as optional.
  5. Turned the Prisma steps into a numbered list instead of a paragraph ending in "Don't forget!!".
  6. Made "finished" checkable. "Run lint and test" became "both pass with zero errors". Before, an agent could run the tests, see failures, and still stop.

One thing for you to fill in: "Don't put business logic in components" probably gets ignored because the agent doesn't know where that logic should go. Replace the TODO with your real directory, such as src/lib/ or server actions in app/**/actions.ts.

Optional: if agents still use npm or skip lint, put it in a hook instead of the doc. A PreToolUse hook can block npm/yarn commands, and a Stop hook can run pnpm lint && pnpm test. I can set either one up with /update-config if you want.

Real run recorded with claude-code / claude-opus-5-5. Output is shown verbatim, unmodified.

Qué hace

Una guía para escribir cualquier documento que consuma un agente. Su premisa es que las mismas pocas palancas hacen que una Skill, un CLAUDE.md o un documento enlazado sean predecibles, porque el agente debe seguir el mismo proceso en cada ejecución.

Cómo funciona

  1. Punteros de contexto: la redacción de la descripción de una Skill o de una línea en AGENTS.md determina cuándo y con qué fiabilidad el agente llega al material que hay detrás.
  2. Las dos cargas: cada añadido cuesta contexto (texto que se carga siempre) o atención humana (recordar qué documentos existen).
  3. Jerarquía de la información: mantén los pasos en el archivo principal, deja la referencia detrás de punteros y evita la dispersión.
  4. Criterios de finalización: cada paso termina con una condición verificable, para que el agente no se detenga antes de tiempo.
  5. Poda: una única fuente de verdad por significado, elimina las líneas que el modelo obedece de todos modos y prefiere indicar el comportamiento objetivo antes que prohibir el no deseado.
  6. Un segundo archivo cubre las decisiones específicas de las Skills: frontmatter, invocación por parte del usuario frente a invocación por parte del modelo y Skills de enrutamiento.

Ideal para

Escribir o afinar una Skill, un CLAUDE.md o un AGENTS.md.

A tener en cuenta

En nuestra prueba, reescribió un CLAUDE.md de ejemplo, explicó cada cambio y dejó un TODO en lugar de inventar un nombre de directorio.

Notas y riesgos

Solo contiene archivos de instrucciones: sin scripts y sin acceso a la red. Cuando le pides que edite un archivo, el agente puede escribir en ese archivo de tu proyecto. El paquete también contiene la licencia MIT original (MIT LICENSE) y agents/openai.yaml (nombre visible para Codex).

SKILL.md original4 archivos en este paquete