I built my first CLI – a tool that scans your
  codebase and generates AGENTS.md files, the
  open standard that AI coding tools (Cursor,
  Copilot, Claude Code) automatically read for
  project context.

  The problem: After a year of learning to code
  and building several Next.js projects, I kept
  manually writing AGENTS.md files to tell AI
  assistants about my components, API routes, and
   patterns. Every. Single. Project. It was
  tedious and would drift out of sync as the
  codebase evolved.

  The solution:

  npx @jpoindexter/agent-smith

  What it does differently:

  Unlike Repomix or Code2Prompt (which pack raw
  files), agent-smith extracts structured
  metadata:

  - Components with props and complexity metrics
  - Hooks with client-only detection
  - API routes with methods and auth status
  - Database models (Prisma/Drizzle) with
  relations
  - Design tokens and pattern detection
  - Import graphs and barrel exports

  It generates "Critical Rules" with wrong/right
  code examples so AI actually follows your
  patterns:

  // WRONG
  <div className="rounded border p-4">...</div>

  // RIGHT
  <Card><CardContent>...</CardContent></Card>

  Output: ~10K tokens of structured context vs
  100K+ tokens of raw code.

  Tech: TypeScript, zero config, works on any
  JS/TS codebase.

  GitHub:
  https://github.com/jpoindexter/agentsmith

  A year ago I couldn't have built this. I spent
  a few hours scratching my own itch, and it
  turns out the problem is pretty common. Curious
   what other metadata would be useful to
  extract, or if there are edge cases I'm
  missing.