review-all-prds
Skill40 starsChanged 7 months agoMITView on GitHub
Holistic cross-PRD review — reads every PRD together to find contradictions, dominant strategies, cognitive overload, broken funnels, and missing-system gaps that single-PRD review cannot detect. Run after all MVP PRDs are individually approved, before architecture begins.
In this repository77 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- design/CLAUDE.mdCLAUDE.md
- docs/CLAUDE.mdCLAUDE.md
- src/CLAUDE.mdCLAUDE.md
- tests/CLAUDE.mdCLAUDE.md
- adoptSkill
- architecture-decisionSkill
- architecture-reviewSkill
- asset-auditSkill
- asset-specSkill
- balance-checkSkill
- brainstormSkill
- bug-reportSkill
- bug-triageSkill
- changelogSkill
- code-reviewSkill
- consistency-checkSkill
- content-auditSkill
- create-architectureSkill
- create-control-manifestSkill
- create-epicsSkill
- create-storiesSkill
- day-one-patchSkill
- design-bibleSkill
- design-reviewSkill
- design-systemSkill
- dev-storySkill
- estimateSkill
- extractSkill
- gate-checkSkill
- helpSkill
- hotfixSkill
- launch-checklistSkill
- localizeSkill
- map-systemsSkill
- milestone-reviewSkill
- onboardSkill
- patch-notesSkill
- perf-profileSkill
- prd-reviewSkill
- project-stage-detectSkill
- propagate-design-changeSkill
- prototypeSkill
- qa-planSkill
- quick-designSkill
- regression-suiteSkill
- release-checklistSkill
- retrospectiveSkill
- reverse-documentSkill
- scope-checkSkill
- security-auditSkill
- setup-frameworkSkill
- skill-improveSkill
- skill-testSkill
- smoke-checkSkill
- soak-testSkill
- sprint-planSkill
- sprint-statusSkill
- startSkill
- story-doneSkill
- The first 60 shown.
The file
Show its rules and text
We split this skill into the rules it gives an agent
16 rules. About 0 are common, 12 are distinctive, and 4 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈12
Distinctive · estimate
This author's own choices. Worth a look.
4
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 14
Read-only. Produces a printed report.
- Very specific: line 21
All v1 PRDs are individually APPROVED
- Very specific: line 22
Before /create-architecture — architecture should not codify contradictions
- Distinctive: line 23
Before any major launch milestone
- Distinctive: line 24
After a batch of PRD changes (3+ PRDs touched in a sprint)
- Very specific: line 44
Verify all MVP-scoped PRDs are individually APPROVED.
- Distinctive: line 46
If any PRD is not yet APPROVED, ask:
- Distinctive: line 50
If user proceeds with degraded run, mark the report header accordingly.
- Distinctive: line 56
Read all PRDs. Hold them in context simultaneously — that is the point of the skill (and why it runs on the opus tier).
- Distinctive: line 90
When is it asked?
- Distinctive: line 123
If the project targets both iOS and Android, scan for:
- Distinctive: line 130
Look at the cross-PRD flow.
- Distinctive: line 210
If verdict is PASS:
- Distinctive: line 213
If CONCERNS or REVISIONS REQUIRED:
- Distinctive: line 230
Every finding names specific PRDs (no "the PRDs say X")
- Distinctive: line 231
Every CRITICAL and HIGH has a concrete fix path
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Review All PRDs (Holistic)
Single-PRD review catches missing sections. This skill catches interaction failures between PRDs that look fine alone but collide once shipped together: the paywall blocks the empty-state CTA, two systems both want primary tab position, the onboarding promises a feature gated by a paywall the user has not seen, the analytics event vocabulary is inconsistent across systems.
Read-only. Produces a printed report. The user runs /propagate-design-change or /design-system retrofit ... to resolve findings.
Purpose / When to Run
Run when:
- All v1 PRDs are individually
APPROVED - Before
/create-architecture— architecture should not codify contradictions - Before any major launch milestone
- After a batch of PRD changes (3+ PRDs touched in a sprint)
Distinct from /consistency-check (value drift) — this skill assesses design intent, flow continuity, and system overlap.
Inputs
- All
design/prd/*.md design/concept.mddesign/systems-index.mddesign/registry/entities.yamldesign/design-bible.mdif present
Outputs
- A printed structured report. No file writes.
Phase 1: Pre-Check
Verify all MVP-scoped PRDs are individually APPROVED. Read design/systems-index.md and confirm Status column.
If any PRD is not yet APPROVED, ask:
- Prompt: "
<N>PRDs are not yet APPROVED. Run holistic review anyway, or stop and fix individual PRDs first?" - Options:
Run anyway (degraded),Stop,Show me which PRDs aren't approved
If user proceeds with degraded run, mark the report header accordingly.
Phase 2: Read Everything
Read all PRDs. Hold them in context simultaneously — that is the point of the skill (and why it runs on the opus tier).
Build cross-PRD summaries:
- Screen-by-screen map — for each screen referenced across PRDs, list every PRD that touches it
- Event vocabulary — every analytics event mentioned, by name, sorted; flag near-duplicates (
paywall_shown,paywallShown,payway_shown) - Configurable values — every value name across PRDs
- Cross-system flows — every flow that crosses 2+ systems (e.g., "user signs up → first session → first paywall trigger")
- Permission asks — every iOS/Android permission requested across PRDs (notifications, camera, location, contacts, photos)
- Background work — every system that schedules background tasks
- Push notifications — every category / channel / trigger across PRDs
- Tabs / nav slots — every system that wants top-level tab placement
Phase 3: Run the Cross-PRD Checks
3a: Funnel breakage
Trace the v1 happy path end to end (typically: install → onboarding → first core action → retention loop → optional monetization). For each handoff between systems, confirm:
- The departing system finishes in a state the receiving system expects
- The receiving system's preconditions are described somewhere
Findings:
- "Onboarding ends with
auth_complete=true, but the home screen PRD assumesprofile_complete=true— there is no system that runs profile setup between."
3b: Surface conflicts
Two PRDs both claim primary visibility:
- "Both Notifications PRD and Activity Feed PRD describe a tab named 'Updates' in primary nav — only one tab can occupy that slot."
- "Onboarding PRD shows a banner on the home screen for 7 days; Promotions PRD shows a banner in the same slot starting day 1."
3c: Permission timing
For each permission:
- When is it asked?
- Is the ask justified by an in-context user action?
- Are multiple PRDs racing to ask the same permission first?
iOS / Android best practice: ask in context, never on first launch unless required for core function.
Findings:
- "Notifications and Location both want to prompt during onboarding. Only one should — pick the highest-leverage one."
3d: Vocabulary inconsistency
Event names with near-duplicates suggest the analytics PRD is not enforced. Configurable values with similar names ("debounce_search_ms" and "search_debounce_ms") suggest no naming convention.
Findings:
- "Analytics event names span three styles: snake_case (38), camelCase (12), Title Case (4). Pick one and standardize."
3e: Dominant-strategy / unintended exploits
Across PRDs, look for:
- Paid features unintentionally accessible through a free path (e.g., share-to-self workaround)
- Onboarding rewards stackable in unintended sequences (account churn for repeated sign-up bonus)
- Empty states with CTAs that lead to a paywall
3f: Cognitive overload at MVP
Count user-visible primary actions, top-level menu items, modal triggers, push categories. If any of these exceed common mobile thresholds, flag:
- More than 5 primary tabs (most apps cap at 4-5)
- More than 3 modals possible on the same screen
- More than 6 push notification categories at MVP
- Onboarding longer than 5 screens at MVP
3g: Cross-platform divergence
If the project targets both iOS and Android, scan for:
- Features described as iOS-only without a corresponding Android note
- Different monetization models per platform (often a real choice, but should be explicit)
- Push categories that map cleanly to iOS but not Android channels (and vice versa)
3h: Missing-system gaps
Look at the cross-PRD flow. Is there a system that should obviously exist that no PRD covers?
- "No PRD describes account deletion flow. Required by iOS App Store and Play Store."
- "No PRD describes data export — required by GDPR if you have EU users."
- "No PRD addresses how to handle a kid who is under 13 — relevant if any user-generated content."
3i: Concept-doc alignment
For each PRD, confirm:
- It serves the JTBD from concept Section 2
- It does not contradict an anti-feature in concept Section 4/5
- It contributes (positively or neutrally) to the primary metric in concept Section 5
Findings:
- "Concept says primary metric is D7 retention; the Onboarding PRD optimizes for D0 conversion (paywall in first session) — these may be in tension. Confirm the trade-off is intentional."
Phase 4: Optional Specialist Synthesis
Spawn Task to:
ux-designer— review the screen-by-screen map for flow coherencepm— review the funnel and metric alignmentaccessibility-specialist— survey accessibility coverage across UI PRDs
Specialists return findings; the main review integrates them.
Phase 5: Verdict
Tally findings:
- CRITICAL — would break the app or fail store review (account deletion missing, IAP receipt invalidation missing, accessibility floor unmet across multiple PRDs)
- HIGH — would harm the user experience or metric (funnel break, surface conflict, permission timing collision)
- MEDIUM — improvable (vocabulary inconsistency, cognitive overload at edges)
- LOW — cosmetic
Verdict ladder:
- 0 CRITICAL, ≤ 3 HIGH → PASS WITH NOTES
- 0 CRITICAL, 4-8 HIGH → CONCERNS
- 1+ CRITICAL or 9+ HIGH → REVISIONS REQUIRED
Phase 6: Output
# Holistic PRD Review
**Verdict: <PASS WITH NOTES / CONCERNS / REVISIONS REQUIRED>**
PRDs reviewed: <N>
Findings: <count by severity>
## CRITICAL
1. <finding> — affecting PRDs: <list>
Fix: <specific recommendation>
## HIGH
1. <finding> — affecting PRDs: <list>
Fix: <specific recommendation>
## MEDIUM
- <list>
## LOW
- <list>
## Strengths to preserve
- <what is working — name PRDs that demonstrate good practice>
## Recommended next steps
- For each CRITICAL: <skill or action>
- For HIGH cluster: run `/propagate-design-change` after each PRD edit
- After fixes: re-run this skill
- If PASS: proceed to `/create-architecture`Phase 7: After-Action
If verdict is PASS:
- Note in output: "Architecture work can proceed. Run
/create-architecturenext."
If CONCERNS or REVISIONS REQUIRED:
- Suggest the order in which to fix: CRITICAL first, then by impact
- Note that
/propagate-design-changeshould follow each PRD edit so dependent PRDs and ADRs are surfaced
Edge Cases
- Single PRD only: skill is wasted — direct user to
/prd-review. - Concept doc absent: degrade — skip Section 3i but run the rest. Note the limitation.
- Mixed maturity (some PRDs polished, others stub): flag the immature ones at the top of the report; do not let them drown out findings about the polished ones.
Quality Gates
- Verdict matches tally — no soft passes when CRITICAL exists
- Every finding names specific PRDs (no "the PRDs say X")
- Every CRITICAL and HIGH has a concrete fix path
- The skill never modifies files
Examples
12 PRDs.
- 1 CRITICAL: no account-deletion PRD (Apple guideline 5.1.1(v))
- 3 HIGH: tab conflict, vocabulary drift (43 events, 9 near-duplicates), funnel break between onboarding and home
- 5 MEDIUM, 2 LOW
- Verdict: REVISIONS REQUIRED
- Recommendation: author missing PRD via
/design-system account-deletion, then re-run the skill.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

