agentprep

Make any repo agent-ready in one command.

agentprep reads your project, learns its real conventions, and writes the AGENTS.md + portable Agent Skills that make AI coding agents — Claude Code, Cursor, Codex, Copilot, Gemini CLI — actually follow your patterns instead of generic ones.

npx agentprep

That's it. Run it in your project root.

---

The problem

You adopted an AI coding agent. It's fast — and it keeps ignoring how your codebase actually works. It puts files in the wrong place, invents a second way to handle errors, skips your tests, reformats lines you didn't touch. So you spend your day re-typing the same context: "we use pnpm, tests are Vitest, run the linter before you finish, components go in src/components…"

The fix the ecosystem landed on is AGENTS.md (project context) plus Agent Skills (reusable, portable capabilities in a SKILL.md format that now works across 20+ tools). The catch: writing good ones by hand, per project, is tedious — so most people don't, or they copy generic templates that don't match their repo.

agentprep writes them from your actual code.

What it does

Point it at a repo. It detects your stack and conventions, then generates:

file	what it is
AGENTS.md	Cross-tool project context: stack, setup, real commands, structure, conventions, working agreements.
CLAUDE.md	A thin pointer so Claude Code picks it up too.
skills/verify-changes/SKILL.md	"Run/test/lint/build the way this project does, before declaring done."
skills/project-conventions/SKILL.md	Coding standards & structure inferred from your code.
skills/write-tests/SKILL.md	How tests are written/run here (only if a test runner is found).

Everything is plain Markdown you own and can edit. Nothing is locked to one tool.

Detects

• Runtimes: Node (npm · pnpm · yarn · bun), Python (pip · poetry · pipenv), Rust, Go, Ruby
• Frameworks: Next.js, React, Vue, Nuxt, SvelteKit, Svelte, Angular, Astro, Remix, Express, Fastify, NestJS, Koa, Django, Flask, FastAPI
• Tooling: Vitest, Jest, Mocha, Playwright, pytest · ESLint, Biome, Ruff · Prettier, Black · TypeScript · monorepos
• Commands: install / dev / build / test / lint / format / typecheck — read straight from your scripts
• Structure: annotates src/, app/, components/, tests/, packages/, and more

Example

Run inside a Next.js + TypeScript app:

$ npx agentprep

  agentprep · make this repo agent-ready
  ~/code/acme-dashboard

  project       acme-dashboard
  stack         Next.js  TypeScript  node/pnpm
  tooling       Vitest, ESLint, Prettier

  ✓ written:
     + AGENTS.md
     + CLAUDE.md
     + skills/verify-changes/SKILL.md
     + skills/project-conventions/SKILL.md
     + skills/write-tests/SKILL.md

The generated AGENTS.md:

# AGENTS.md

## Project
Internal analytics dashboard for Acme.

**Stack:** Next.js · TypeScript · node (pnpm)

## Setup & commands
- Install: `pnpm install`
- Dev:     `pnpm run dev`
- Test:    `pnpm run test`
- Lint:    `pnpm run lint`

## Conventions
- Written in TypeScript — prefer typed code; avoid `any`.
- Code is formatted with Prettier — match existing style, don't reformat unrelated lines.
- Tests use Vitest — add/adjust tests for behavioural changes.
…

Usage

npx agentprep            # scan ./ and write the files
npx agentprep ./my-app   # scan a specific directory
npx agentprep --print    # preview everything, write nothing

Skills are portable, but each tool looks in its own folder. Symlink once:

ln -s ../../skills .claude/skills    # Claude Code
ln -s ../skills .agents/skills       # Codex & others

Then commit. Your agents now read your conventions on every task.

Honest limitations

agentprep gives you a strong, accurate starting point — not a finished spec. It infers from manifests and structure, so:

• It can't know rules that live only in your head (domain logic, review etiquette, release process). The files are made to be edited — that's the workflow.
• Framework/tool detection covers the common cases above; exotic setups fall back to sensible defaults.
• A --ai mode (use an LLM to infer deeper conventions from the code itself) is on the roadmap — see issues.

Contributing

Want it to detect your stack (Laravel? Spring? Deno? Elixir?) or generate a skill you rely on? Open an issue or PR — detectors are small and self-contained.

License

MIT

_{Built by @elonmust26 · for the agent-coding era.}