Skill: agents-generator

[!WARNING] [Authorized Use Only] This skill writes or updates AGENTS.md, .agents/rules/, optional platform instruction files, and timestamped backups in the target project. Read the detected inputs and proposed outputs first, obtain approval before changing target files, and use it only inside the user's intended project scope.

When to Use

Use this skill when the user wants to:

• create a complete, project-specific AGENTS.md instead of generic agent rules;
• generate companion rules for detected frameworks, tests, databases, styling, or monorepo packages;
• create a minimal AGENTS.md, preview changes without writing, 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 scope, or to treat generated guidance as a substitute for human review.

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

User says	Mode	Output
"simple AGENTS.md", "just the basics", "minimal"	Minimal	Single AGENTS.md (~30 lines, no rule files)
"full AGENTS.md", "with rules", "complete", or default	Full	AGENTS.md + .agents/rules/*.md
"update AGENTS.md", "refresh", "my stack changed"	Update	Diff existing, regenerate only what changed

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, but never read secrets. Read package.json, non-secret config files, and directory structure before generating anything. Never open .env, .env.local, credential stores, or similarly secret-bearing files. Derive environment variable names only from .env.example placeholders and source references such as process.env.NAME, without reading or reporting values.
• 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.
• Do not execute project scripts by default. Package-manager scripts are repository-controlled shell entry points. Detect and document candidate format/lint commands, but do not run them unless the user separately requests execution after the exact script body and invoked tooling have been reviewed.
• 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 non-secret config files and explore directory structure. Exclude .env* files other than placeholder-only .env.example; never read secret values.
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

• Report the detected [format cmd] and [lint cmd] as unexecuted candidates. Run neither automatically; execute one only after the user separately authorizes it and its exact project-controlled script body has been reviewed.
• 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

Limitations

• Generated instructions are proposals and require human review before they are adopted or committed.
• Command validation is limited to scripts and files visible in the target project; it cannot prove that tools, services, or platform-specific commands will work in every environment.
• Project-provided package scripts are untrusted executable code. Generation and documentation of a script do not authorize running it.
• The skill does not authorize writes outside the intended project scope or replace project-specific security, build, or deployment review.

References

Priority	File	Purpose
Required	assets/agents-full.md	Full AGENTS.md template with all 25+ sections and filling rules
Required	assets/agents-minimal.md	30-line agents.md standard template
Full mode	assets/architecture.md	Architecture rules template
Full mode	assets/frontend-patterns.md	Frontend patterns template
Full mode	assets/server-actions.md	Server actions / backend template
Full mode	assets/testing.md	Testing strategy template
Full mode	assets/git-workflow.md	Git workflow template
Full mode	assets/sdd-workflow.md	SDD workflow template
Full mode	assets/styling.md	Styling rules template
Full mode	assets/forms.md	Form patterns template
Full mode	assets/database.md	Database rules template
Full mode	assets/i18n.md	i18n rules template
Full mode	assets/backend.md	Backend/NestJS template
Conditional	assets/claude.md	CLAUDE.md — only if Claude detected
Conditional	assets/platform.md	Multi-platform files
Conditional	assets/agents-nested.md	Monorepo nested AGENTS.md
Reference	references/decision-matrix.md	Full detection logic and edge cases
Reference	references/example-output/README.md	Quality benchmark
Reference	references/template-filling-guide.md	Placeholder filling rules