Live test results — not marketing

What Prompt Modules actually does

We ran a full automated test suite: 3 domains, 9 real bugs fixed, 4 module versions. Here's every number.

2026-04-14

Test run

41.8s

Duration

Modules tested

9/9

Fixes applied

100%

Success rate

The @reuse effect

The first time you define a module, PM builds a rich PML structure (~450 tokens). Every subsequent use is just @reuse[module_name] — 5 tokens.
That's where the real saving lives.

Without PM (full context every time)

With PM (@reuse after first definition)

tokens per @reuse

vs ~500 without PM

98%

token reduction at 100 uses

49,067 tokens saved

$0.74

saved per module at 100 uses

at Claude API pricing

Module quality: B → A after 3 fixes

Every fix Claude applies gets versioned into the module. After 3 real bugs fixed, all 3 modules jumped from Grade B (72%) to Grade A (82%) — production-ready.

user_managementauthGrade A

v4 · 3 fixes applied

82%

+10% improvement

Before fixes: 72%After fixes: 82%

Bugs fixed and versioned:

✓Google OAuth callback fails on mobile Safari → v2

✓JWT tokens not invalidated on logout → v3

✓Role check missing on /api/admin routes → v4

@attention rules (AI guardrails):

never_store_passwords_plaintextalways_validate_jwt_server_sidemust_check_rls_on_all_tablesnever_expose_service_role_key

payment_escrowpaymentsGrade A

v4 · 3 fixes applied

82%

+10% improvement

Before fixes: 72%After fixes: 82%

Bugs fixed and versioned:

✓Duplicate webhook processing causing double upgrades → v2

✓Trial expiry not enforced after payment fails → v3

✓Subscription cancel not reflected immediately → v4

@attention rules (AI guardrails):

never_process_payments_without_idempotency_keyalways_verify_webhook_signaturemust_enforce_trial_expiry_server_sidenever_store_card_data

rag_systemaiGrade A

v4 · 3 fixes applied

82%

+10% improvement

Before fixes: 72%After fixes: 82%

Bugs fixed and versioned:

✓Chunk overlap missing — RAG misses cross-boundary context → v2

✓Embeddings not normalized — wrong cosine similarity → v3

✓No re-ranking step — top-k not always most relevant → v4

@attention rules (AI guardrails):

never_mix_user_document_spacesalways_normalize_embeddings_before_storagemust_include_source_citations_in_every_answernever_retrieve_from_other_users_documents

The fix system: Claude learns, permanently

When a bug is found, pm fix records the issue + solution and creates a new module version. Next session, Claude already knows the fix.

Module evolution — user_management

v1CreatedAuto-analyzed from raw prompt · @core + @attention generated

v2Fix #1OAuth callback mobile Safari → redirect_uri must use https://

v3Fix #2JWT blacklist added → tokens now invalidated on logout

v4Fix #3requireRole("admin") added → server-side admin check enforced

9/9

fixes applied successfully

3.0

avg fixes per module

+10%

quality score per 3 fixes

Cross-prompt reuse: 7 modules auto-detected

When we ran a complex B2B SaaS prompt against the registry, PM found 7 matching modules automatically — no manual tagging needed.

Claude Code — auto-detect output

# Complex prompt: B2B SaaS with auth, payments, RAG, admin, team workspaces

# Registry scan → 7 matches found:

↺ @reuse[user_management]→ injected

↺ @reuse[payment_escrow]→ injected

↺ @reuse[rag_system]→ injected

↺ @reuse[ai_engine]→ injected

↺ @reuse[file_storage]→ injected

↺ @reuse[admin_panel]→ injected

↺ @reuse[api_gateway]→ injected

# Result: 126 tokens → 5 tokens per @reuse tag × 7 = 35 tokens total context

What makes a module production-ready

We score every module on 10 dimensions. Here's what each point means for Claude's output quality.

@module declaration

5 pts

Named identity — Claude knows exactly what this is

@core blocks (×5)

15 pts

Structured knowledge — implementation details per domain

@attention rules

10 pts

AI guardrails — what Claude must never/always do

Token efficiency

10 pts

Under 400 tokens = fast injection, low cost

@version marker

5 pts

Versioned — Claude knows what changed

Explicit constraints

10 pts

never_X / always_X / must_X boundaries

Fix iterations

15 pts

Battle-tested — real bugs found and fixed

Naming (snake_case)

5 pts

Consistent — works across all projects

Metadata / tags

5 pts

Searchable and categorized

Structure depth

5 pts

Nested blocks = richer context

80–100%

Production ready

65–79%

Good, needs polish

50–64%

Functional

0–49%

Needs work

Cumulative savings over time

One module, one domain. Tokens and dollars saved as usage grows. (Assumes 500 tokens of context without PM, 5 tokens with @reuse after first definition)

Uses	Without PM	With @reuse	Tokens saved	$ saved	Savings %
1×	500	438	+62	$0.0009	12%
5×	2,500	458	+2,042	$0.0306	82%
10×	5,000	483	+4,517	$0.0678	90%
25×	12,500	558	+11,942	$0.1791	96%
50×	25,000	683	+24,317	$0.3648	97%
100×	50,000	933	+49,067	$0.7360	98%

Start saving tokens today

Your modules. Your registry. Your savings.

Free plan includes 5 modules. No credit card. The test suite code is open — run it yourself.

Get started free →Browse modules

Test run: 2026-04-14 · 41.8s · 100% automated · no cherry-picking