Agent Rules Generator

The Problem

Every new project starts the same way: your AI agent has zero context. It doesn't know you use App Router instead of Pages Router. It doesn't know your auth code uses JWT with a custom refresh flow, not session cookies. It doesn't know the path alias is @/, not ../../../. So you spend the first 15-30 minutes "teaching" it conventions that should already be in its config files — and by session 4, it's forgotten half of them anyway.

Meanwhile, you're switching between Cursor on one project and Claude Code on another, maintaining separate rule files that drift apart. A Next.js 15 upgrade breaks your old rules. Your team lead writes a CLAUDE.md and nobody reads past line 200. The config becomes stale documentation nobody trusts.

What You Get

• Automated project discovery — detects tech stack, frameworks, build systems, and existing config files from project structure before generating any rules
• Framework-specific rule templates — React+Next.js (App Router conventions, server components), Python+Django (service layer, migrations), TypeScript (strict mode, brand types), Go (module structure), Rust (cargo conventions)
• Progressive disclosure architecture — safety rules in first 20 lines, architecture in 20-40, conventions in 40-70, nice-to-have beyond that; rules the agent actually processes first
• Cross-tool rule translation — migrate .cursorrules JSON to CLAUDE.md markdown and back, preserving intent while adapting format (file-scoped .mdc rules ↔ section-organized CLAUDE.md)
• Mono-repo multi-service support — root CLAUDE.md for shared conventions + per-service configs for stack-specific patterns across Go, Python, TypeScript services
• Config drift detection — automated check comparing rule references against actual project structure; flags when a framework upgrade makes old rules give wrong advice

Why This Beats Prompting It Yourself

Writing "follow React best practices" is useless — the agent already knows React. The skill captures what's specific to your project that differs from framework defaults. Instead of a 300-line reference document the agent skims, you get a 50-line operating manual where every rule answers: "What would a senior dev who's been on this project for 6 months know that isn't obvious from the code?" Rules are phrased as actionable directives with trigger conditions, not vague suggestions.

Example Session

You: "Set up agent config for this Next.js + Prisma + Tailwind project. I use both Cursor and Claude Code."

Agent:

# Project Discovery Complete
tech_stack: [nextjs-app-router, typescript, tailwindcss, prisma]
build_system: [npm, github-actions]
test_framework: vitest
existing_config: none

# Generated Files

## CLAUDE.md (42 lines)
**Safety Rules (lines 1-12):**
- Never modify prisma/migrations/ — they are auto-generated
- Never modify node_modules/ or .next/ — rebuild instead

**Architecture (lines 13-28):**
- Next.js App Router (app/ directory). Server components by default.
- Data fetching via server components + Prisma ORM, no useEffect for data
- Route handlers in app/api/*/route.ts (one per HTTP method)

**Conventions (lines 29-42):**
- Path alias: @/ maps to src/
- Tailwind only — no CSS modules, no styled-components
- Components: PascalCase, named exports, one per file
- Server actions: "use server" at file top or in dedicated action files

## .cursor/rules/ (3 .mdc files)
- security-rules.md

Use Cases

• New project bootstrap — configure agent context in 30 seconds when cloning or creating a project
• Tool migration — translate 15 Cursor .mdc files into a single CLAUDE.md when switching to Claude Code
• Team standardization — generate consistent agent configs across a mono-repo with 8 microservices in different stacks
• Framework upgrade — re-generate rules after a major version bump that changes APIs and conventions
• Rule bloat reduction — take a 400-line CLAUDE.md that the agent ignores and restructure it into a focused, progressive-disclosure format

Known Limitations

This skill generates config files but cannot guarantee 100% agent compliance — some failures are model capability issues, not config quality issues. Framework-specific templates cover the most common stacks (React, Next.js, Django, FastAPI, Go, Rust, TypeScript); highly custom or niche frameworks may need manual rule additions.