Generate project-specific AGENTS.md and rule files

This skill generates a tailored, project-specific AGENTS.md plus companion .agents/rules/*.md files for a target project - not a placeholder template, but a living document derived from the project's real toolchain, detected package manager, frameworks, test setup, and directory structure.

Use it to create a complete, project-specific AGENTS.md instead of generic agent rules; to generate companion rules for detected frameworks, tests, databases, styling, or monorepo packages; to create a minimal AGENTS.md, preview changes without writing (dry-run), or update existing instructions after the stack changes. Do not use it to invent conventions without inspecting the target project, to overwrite instructions outside the user's intended scope, or to treat generated guidance as a substitute for human review - it is authorized-use only, requiring approval before changing target files.

Input is the target project's package.json, lockfile (used to detect the package manager - bun.lock, pnpm-lock.yaml, package-lock.json, or yarn.lock, never defaulting to npm), non-secret config files, and directory structure; .env* files other than a placeholder-only .env.example are never read, and secret values are never reported. Three modes select the output: Minimal (a single roughly 30-line AGENTS.md, triggered by phrases like "simple" or "just the basics"), Full (AGENTS.md plus .agents/rules/*.md, the default), and Update (diff the existing files against re-detected project state and regenerate only what changed). A dry-run mode runs detection without writing, showing a detection summary, files that would be created, skipped rules, and sample output. Rules are only generated for categories that apply - a frontend-only project skips backend and database rules, for example. Hard rules govern generation: validate that every output command exists as a script key in package.json; scan output for placeholder markers ({{, TODO, add here, ...) and reject if any remain; back up existing files to .agents/backups/ with a timestamp before overwriting; and never execute a project's own scripts by default, since package-manager scripts are untrusted repository-controlled code - detected format and lint commands are reported as candidates, not run. If a Claude-specific setup is detected, a thin CLAUDE.md is also generated; multi-platform instruction files and nested monorepo AGENTS.md files are generated when applicable. Output always reports the mode used and why, files created or modified, a full detection summary, which rules were generated versus skipped (with reason), and a confidence score.

The skill detects and adapts to whichever package manager, framework, test runner, and monorepo layout the project actually uses. The source's own worked example: a Bun, Next.js, Tailwind, Vitest, and Server Actions project gets an AGENTS.md plus six rule files - architecture, frontend patterns, server actions, testing, git workflow, and SDD workflow - while backend, database, i18n, and forms rules are skipped for that stack because there is no NestJS, ORM, hardcoded-locale need, or form library present. If the generated AGENTS.md would exceed 300 lines it warns, and above 500 lines it moves content into rule files instead.

Anyone setting up or maintaining agent instructions for a real codebase who wants AGENTS.md and its companion rules to reflect the project's actual toolchain rather than a generic template, with human review still required before the generated guidance is adopted or committed.