---

name: generate-ai-rules description: Generate AI assistant configuration files for a repository — CLAUDE.md, AGENTS.md, and Cursor rules (.cursor/rules/*.mdc) — from codebase analysis. Use whenever the user wants to create or update CLAUDE.md, AGENTS.md, agent rules, Cursor rules, AI coding assistant configuration, or "onboard AI tools" to a project, even if they only mention one of the file types.

Generate AI Rules

Generate configuration files that help AI coding assistants work effectively with a codebase. Three targets, generated from the same analysis so they stay consistent:

1. AGENTS.md — the cross-tool standard (agents.md), read by most AI coding tools including Claude Code, Cursor, Codex, and Gemini CLI.
2. CLAUDE.md — Claude Code's project instructions file.
3. .cursor/rules/*.mdc — Cursor's scoped project rules.

Workflow

1. Determine targets and gather data

• Generate all three targets by default; the user may skip any (e.g., "skip cursor rules", "keep my existing CLAUDE.md").
• If a target file already exists and the user didn't say to regenerate it, ask whether to update it or leave it alone.
• Check <repo>/.ai/docs/ for analysis documents from the analyze-codebase skill. If present, use them as the primary source (spot-check against the code — they may be stale). If absent, offer to run analyze-codebase first, or explore the codebase directly for a quicker pass.

2. Generate the files

Shared principles for all targets:

• Accuracy: every command, path, and convention must come from the actual project. Test that commands at least look right against the manifest files (e.g., scripts in package.json, tasks in Makefile, uv run vs pip).
• Actionability: specific, executable instructions beat vague guidance. "Run uv run ruff format src/" beats "format your code".
• Conciseness: these files are loaded into every AI session — every line costs context. Only include what the AI cannot cheaply discover by reading the code: commands, non-obvious conventions, gotchas, things that have gone wrong before. Do not restate what the code structure makes obvious.
• Consistency: same terminology and architecture descriptions across all generated files.

AGENTS.md

The primary file — write it first, and write it best. Target well under 150 lines.

• Project overview (1–2 sentences)
• Build, test, run, lint commands (in backticks, copy-pasteable)
• Architecture overview (3–5 bullets)
• Code style conventions
• Testing instructions
• Git workflow (commit format, PR process)
• Key project-specific conventions and gotchas

CLAUDE.md

Claude Code reads AGENTS.md natively, so avoid duplicating content between the two files. Pick based on what exists and what the user wants:

• If AGENTS.md is generated/present (recommended): make CLAUDE.md a thin complement — a single line See AGENTS.md for project instructions. plus only Claude-specific additions if any (e.g., skill/subagent usage preferences, permission notes). If there is nothing Claude-specific, ask the user whether they want CLAUDE.md at all.
• If the user wants a standalone CLAUDE.md (no AGENTS.md): include the full content — overview, commands, style, architecture, key components, gotchas, known issues. Target under 300 lines; long CLAUDE.md files degrade rather than improve AI performance.

.cursor/rules/*.mdc

Generate 2–3 focused, composable rule files in MDC format (markdown with YAML frontmatter):

---
description: Brief description of what this rule covers
globs:
  - "src/**/*.py"
alwaysApply: false
---

# Rule Title

Content...

• project-overview.mdc — project context, architecture, conventions (alwaysApply: true, no globs needed)
• code-patterns.mdc — code style, testing patterns, anti-patterns to avoid (globbed to source files)
• api-conventions.mdc — only if the project has a significant API surface (globbed to API/handler files)

Keep each file to 50–100 lines. Rules should be prescriptive and project-specific, with short code examples from the actual codebase. Reference files with @path syntax where helpful. If a legacy .cursorrules file exists, migrate its still-valid content into the new files and tell the user the legacy file can be removed.

3. When existing files are provided

When updating rather than creating:

• Preserve the existing structure, tone, and any manually added sections not derivable from analysis (they usually encode hard-won knowledge).
• Refresh outdated information: stale commands, renamed paths, removed components.
• Tell the user specifically what you changed and why.

4. Report

List the files written, their line counts, and anything you left out or couldn't verify. If the repo's docs and reality diverged notably, mention it — that's a signal the team should know.