Consistency Check

Compares what each PRD claims about shared entities, values, and formulas against the registry. Pure read — never writes a file. Produces a structured findings list the user can resolve with /design-system retrofit ... or /propagate-design-change.

Purpose / When to Run

Run when:

• Three or more PRDs exist
• A PRD has just been authored that overlaps with prior PRDs
• Before /review-all-prds (which is more expensive)

Distinct from /review-all-prds: this skill checks values for drift; the holistic review checks design intent for contradictions. Run consistency-check first — it is much cheaper.

Inputs

• design/registry/entities.yaml
• All design/prd/*.md
• Optional: design/systems-index.md

Outputs

• Printed findings. No file writes.

Phase 1: Load the Registry

Read design/registry/entities.yaml. The registry is the source of truth for cross-PRD facts. Each entry has:

• name
• type (entity / configurable / formula / constant / event / endpoint)
• value or attributes
• source (the PRD that owns the canonical definition)
• referenced_by (other PRDs that reference it)

If the registry does not exist, stop:

"No design/registry/entities.yaml found. The registry is built incrementally as /design-system runs. Cannot consistency-check without it. Either author more PRDs or run /design-system retrofit ... to populate the registry from existing PRDs."

If the registry has fewer than 5 entries, warn that meaningful checks need ≥5 entries — the skill will run but findings will be sparse.

Phase 2: Index Every PRD

For each design/prd/*.md:

1. Read the file.
2. Identify which registry entries it references — grep for each entry's name in the file body.
3. For each referenced entry, extract the value the PRD claims (from sections 3 Detailed Requirements, 6 Configurable Values, and any inline tables).

Build a per-entry table:

entry: rate_limit_login_attempts
  registry value: 5 (per minute)
  source: design/prd/auth.md
  referenced in:
    design/prd/auth.md → 5 per minute ✓
    design/prd/account-recovery.md → 3 per minute ✗ MISMATCH
    design/prd/security.md → "5 attempts" ✓

Use grep judiciously — read each PRD once, scan for known entry names from the registry, capture surrounding context.

Phase 3: Classify Findings

For each entry referenced in 2+ PRDs:

• MATCH — all PRDs agree with the registry. No finding.
• MINOR DRIFT — same value but different units or notation (e.g., "5 per minute" vs. "5/min"). Cosmetic. Reportable.
• VALUE MISMATCH — PRDs claim different values. HIGH finding.
• REGISTRY DRIFT — registry value differs from the source PRD's current text. The PRD changed without registry update. HIGH finding.
• ORPHAN REFERENCE — a PRD references something that looks like a registry entry name but is not registered. MEDIUM finding (may be a typo, may need registration).

Also detect cross-PRD-only patterns (registry-independent):

• NAME COLLISION — two PRDs define different things with the same name. HIGH finding.
• INTERFACE DISAGREEMENT — PRD A says it emits event foo, PRD B says it emits event Foo (case mismatch) or with different payload shape. HIGH finding.
• TIMING DISAGREEMENT — PRD A says "show paywall after 3 sessions", PRD B says "after 5". HIGH if neither references the registry.

Phase 4: Print the Findings

Output format:

# Consistency Check

PRDs scanned: <N>
Registry entries: <N>
Findings: <count by severity>

## HIGH

### 1. VALUE MISMATCH — `rate_limit_login_attempts`
- Registry says: `5 per minute` (source: design/prd/auth.md)
- design/prd/auth.md: `5 per minute` ✓
- design/prd/account-recovery.md: `3 per minute` ✗

Resolution suggestions:
- If 5 is correct → update `design/prd/account-recovery.md` Section 3
- If 3 is correct → update registry source PRD and run `/propagate-design-change`

### 2. INTERFACE DISAGREEMENT — analytics event `paywall_shown`
- design/prd/paywall.md: emits with `{ trigger: string, user_segment: string }`
- design/prd/analytics.md: expects `{ trigger: string, segment: string, ts: number }`

Field name mismatch (`user_segment` vs `segment`) and missing `ts`.

## MEDIUM

### 3. ORPHAN REFERENCE — `feature_flag_paywall_v2`
- Referenced in: design/prd/paywall.md, design/prd/onboarding.md
- Not registered.

Either register it (run `/design-system retrofit design/prd/paywall.md` to push to registry) or rename to a registered flag.

## MINOR

### 4. UNIT DRIFT — `cache_ttl_user_profile`
- design/prd/profile.md: `300 seconds`
- design/prd/auth.md: `5 minutes`
- Same value, different notation. Standardize.

## Verdict
[CLEAN | DRIFT FOUND] — <count> HIGH, <count> MEDIUM, <count> MINOR.

If zero HIGH findings, the verdict is CLEAN. Otherwise DRIFT FOUND.

Phase 5: Recommend Resolution Path

Below the findings list, print a short next-action block:

## Recommended next actions

For each HIGH finding:
- Decide on the canonical value
- Update the non-canonical PRD using `/design-system retrofit <path>` or manual Edit
- Run `/propagate-design-change <changed-prd>` to surface downstream impact
- Re-run `/consistency-check` to verify resolution

After all HIGH are resolved, run `/review-all-prds` for the holistic design check.

Do not run any of these — the user decides.

Edge Cases

• Registry is empty: stop. The skill needs at least one entry to check anything.
• Two registry entries with the same name: that is itself a HIGH finding — the registry is corrupt. Surface it.
• PRD references a registry entry inside a code block or example: still count as a reference but lower confidence. Flag as MINOR if the only reference.
• Whitespace or punctuation differences: do not flag. Compare semantic values, not string equality.

Quality Gates

• Every HIGH finding includes: registry path, conflicting PRD paths, resolution suggestions.
• The verdict is unambiguously CLEAN or DRIFT FOUND.
• No false positives on cosmetic differences (units, capitalization in prose).
• Read-only — no file writes ever.

Examples

5 PRDs, registry has 23 entries.

• Output: 1 HIGH (rate-limit mismatch), 2 MEDIUM (orphan references), 3 MINOR (unit drift). Verdict: DRIFT FOUND.

12 PRDs, registry has 60 entries.

• Output: 0 findings. Verdict: CLEAN.

3 PRDs, all reference an unregistered analytics event:

• Output: 3 MEDIUM (each PRD with the orphan reference). Recommend running /design-system retrofit design/prd/analytics.md to register the event.

Constraints

• Read-only — never writes
• Cross-document only — does not check intra-document consistency (that is /prd-review's job)
• Does not evaluate design quality — only value-level drift