Agents Generator

Skill: agents-generator

Generates a tailored AGENTS.md + .agents/rules/*.md for the target project — not a template with placeholders, but a living document that matches the project's real toolchain.

What you get

From a project that uses Bun + Next.js 16 + Tailwind + Vitest + Server Actions, the skill produces:

AGENTS.md
├── Setup commands: bun install, bun dev, bun run test:run, bun doctor
├── Verification Cycle: bunx tsc --noEmit → bun run lint → bun run test:run → bun doctor
├── Conventions: "Bun always. Plain TypeScript types + guards."
└── Architecture → .agents/rules/architecture.md

.agents/rules/
├── architecture.md       ← ASCII diagram with real directories, exact versions
├── frontend-patterns.md  ← Component rules, state locations, trust boundaries
├── server-actions.md     ← downloadVideo() flow, DownloadResult type, rate limiter
├── testing.md            ← "74 tests in 5 files", vitest commands, mock patterns
├── git-workflow.md       ← Conventional commits, pre-commit checks
└── sdd-workflow.md       ← Preflight defaults, post-apply verification

Rules NOT generated: backend.md (no NestJS), database.md (no ORM), i18n.md (hardcoded Spanish), forms.md (manual inputs), styling.md (Tailwind in frontend rules).

Activation Contract

Generate AGENTS.md + .agents/rules/*.md for the target project. Never guess — read the project's actual files first.

Mode selection

Dry-run mode

If the user asks to "preview", "show what would change", "dry-run": run all detection but do NOT write files. Show detection summary, files that would be created, skipped rules, and sample output.

Hard Rules

• Read before writing. Read package.json, all config files, and directory structure before generating anything.
• Detect package manager FIRST. Check lockfiles: bun.lock→bun, pnpm-lock.yaml→pnpm, package-lock.json→npm, yarn.lock→yarn. NEVER default to npm. Every command uses the detected PM.
• Generate only what applies. No backend rules for frontend-only. No database rules without ORM.
• Auto-format and lint generated files. Run [format cmd] and [lint cmd] on generated files only — never the whole project. These are fast, safe operations.
• Validate commands. Every command in output must exist as a script key in package.json.
• No placeholders. Scan output for {{, TODO, add here, .... Reject if any remain.
• Backup first. If files exist, copy to .agents/backups/ with timestamp.

Execution Steps

Common

1. git rev-parse --show-toplevel → project root.
2. Detect package manager FIRST: check lockfiles. bun.lock→bun, pnpm-lock.yaml→pnpm, package-lock.json→npm, yarn.lock→yarn. Never default to npm.
3. Read package.json (scripts, deps, workspaces). Save scripts for validation.
4. Read config files and explore directory structure.
5. Select mode (ask if ambiguous).

Full mode

1. Read assets/agents-full.md — this is the AGENTS.md structure with all sections and filling rules.
2. Read project files and fill every placeholder with real data. Never use generic text.
3. Generate AGENTS.md at project root. Wrap content in <!-- AGENTS-GENERATED-START --> / <!-- AGENTS-GENERATED-END -->.
4. For each applicable rule category, read the corresponding template from assets/ and generate the rule file in .agents/rules/.
5. If Claude detected (.claude/ or CLAUDE.md): generate thin CLAUDE.md from assets/claude.md.
6. If platform files detected: generate from assets/platform.md.

Minimal mode

1. Read assets/agents-minimal.md — 30-line agents.md standard format.
2. Generate single AGENTS.md.

Update mode

1. Backup existing files.
2. Re-detect project state.
3. Diff old vs new. Regenerate only changed categories.

Post-generation

• Run [format cmd] on the generated files only.
• Run [lint cmd] on the generated files only.
• Scan for {{, TODO, .... Fix any found.
• Verify all commands exist in package.json scripts.
• If AGENTS.md > 300 lines, warn. If > 500, move content to rule files.
• Summarize all changes using conventional commit format before declaring done.
• Report: what was detected, generated, skipped, and confidence score.

Output Contract

Return:

• Mode used and why
• Files created/modified
• Detection summary (all categories)
• Rules generated and skipped (with reason)
• Confidence score

References