Tokens. Money. Code that doesn't break.
Full automated test suite · 3 modules · 9 bugs fixed · real API calls. Every number below is verifiable.
Token savings
The first prompt defines the module (~450 tokens). Every subsequent call is @reuse[name] — 5 tokens. That's where the saving compounds.
| Uses | Without PM | With @reuse | Saved | % |
|---|---|---|---|---|
| 1× | 500 | 455 | +45 | 9% |
| 5× | 2,500 | 475 | +2,025 | 81% |
| 10× | 5,000 | 500 | +4,500 | 90% |
| 25× | 12,500 | 575 | +11,925 | 95% |
| 50× | 25,000 | 700 | +24,300 | 97% |
| 100× | 50,000 | 950 | +49,050 | 98% |
Real money saved per month
10 modules · 50 prompts/day per module · 22 workdays. That's 5.4M tokens saved/month — before the registry even grows.
Robustness: bugs that never reach production
Every module has @attention rules — hard constraints Claude follows on every generation. No reminders. No prompting.
# Prompt: "implement user login"
// Generated code:
const password = req.body.password
await db.users.create({ email, password }) // ← plaintext!
// JWT validated client-side → security hole
// No RLS on users table → data leak
// Service role key in .env.local → exposed
# Same prompt + user_management v4 injected
// @attention rules active:
// never_store_passwords_plaintext ✓
await supabase.auth.signUp({ email, password })
// always_validate_jwt_server_side ✓
const user = await getUser(req) // server only
// must_check_rls_on_all_tables ✓ // enforced
Fix system: a bug fixed once is fixed forever
When you find a bug, pm fix records it and creates a new module version. Every future session — and every teammate — starts with that knowledge already baked in.
@fuse: modules absorb new stacks automatically
When a new prompt uses a different tech stack for the same module category, PM doesn't replace — it merges. The module gets richer with every project.
@core[authentication] {
provider: supabase_auth
strategies: [email, google_oauth]
}
@core[authentication] {
provider: supabase_auth ← intact
strategies: [email, google_oauth]
}
@core[authentication_clerk] {
provider: clerk ← new stack
middleware: clerkMiddleware()
session: clerk_sessions
}
AI-generated PML: custom to your stack
PM doesn't use generic templates. Claude reads your prompt and generates a module that reflects your actual stack — Clerk, Prisma, LemonSqueezy, whatever you're using.
@core[authentication] {
provider: supabase_auth ← assumes Supabase
strategies: [email, google_oauth]
}
@core[authentication] {
provider: clerk ← from your prompt
middleware: clerkMiddleware()
session: clerk_sessions
framework: nextjs_app_router
}
Module quality: B → A after 3 fixes
Every module is scored on 10 dimensions. 3 real bugs fixed = Grade A. Production-ready.
With PM vs without PM
What changes in your day-to-day development workflow.
| Situation | Without PM | With PM |
|---|---|---|
| New project setup | Explain full stack to Claude every time | pml: → 7 modules extracted in 3 seconds |
| Implementing a feature | Claude guesses your architecture | @reuse injects exact context automatically |
| Context per prompt | ~500 tokens of repeated context | 5 tokens per @reuse |
| Bug appears in production | Fix it, forget it, it comes back | pm fix → versioned → never repeats |
| Switching stack (Supabase → Clerk) | Start module from scratch | @fuse → Clerk added inside the module |
| New teammate onboards | They repeat all your past mistakes | They inherit the full registry instantly |
| Month 6 of a project | Same token cost as month 1 | Registry has 20+ modules, costs stay flat |
| Security guardrails | Claude forgets to check RLS, JWT, etc. | @attention rules enforced on every prompt |
Your registry. Your savings. Forever.
Free plan: 5 modules. No credit card. The test suite is open — run it yourself.
Test: 2026-04-14 · 41.8s · 100% automated · zero cherry-picking