ramsbaby/jarvis/plan-review

/plan-review — 11-Section Rigorous Design Review

Reviews a feature plan, RFC, or architecture doc across 11 independent dimensions. Goal: push the plan to "best-in-class" — not to rubber-stamp it.

When to invoke

• After writing a feature design doc or RFC
• "Is this approach OK?" / "What did I miss?" / "Is there a better way?"
• Before kicking off any nontrivial implementation

11 sections

1. Problem definition — Is the problem stated precisely? What evidence supports it being worth solving?
2. Scope boundaries — What is explicitly in scope, out of scope, and deferred?
3. Architecture — Data flow, dependencies, blast radius of failure
4. Security — Authentication, authorization, secrets handling, PII boundaries, threat model
5. Observability — Logs, metrics, alerts, on-call playbook
6. Deployment — Rollback strategy, gradual rollout, feature flags, dark launches
7. Performance — Latency, throughput, cost budget, scaling assumptions
8. Reliability — Failure modes, recovery paths, SLO targets
9. Testing — Unit, integration, regression, chaos, load
10. Maintainability — Code patterns, documentation, onboarding cost for new contributors
11. Migration — Backward compatibility, data migration, rollout/rollback ordering

Output format

For each section, the review returns:

• Pass — meets bar
• Risk — works but has identifiable risk → suggested mitigation
• Block — critical gap → must address before implementation

Recommended environment

For deep analysis, run with --effort xhigh (Opus 4.7 only) to maximize reasoning depth:

claude --effort xhigh
> /plan-review <plan-doc-path>

Cost approximately 2x baseline, but materially improves gap detection — especially in security and reliability sections.

Self-check (before issuing review — BLOCKING)

1. Did I read the actual plan document (not assume from title)?
2. For each section, do I cite specific lines or claims, not generic feedback?
3. For each Block, do I have a concrete falsification — what evidence shows this would actually fail?
4. Did I avoid groupthink with the plan author (contrarian challenge for at least 3 decisions)?

Limitations

• Cannot detect domain-specific failure modes the author understands but didn't document — surface this gap by asking "what would an expert in domain X say?"
• Cannot replace formal threat modeling for security-critical systems
• 11 sections is a checklist, not a substitute for judgment

Source attribution

Adapted from gstack (paretofilm/superpowers-gstack) /plan-ceo-review. Section list refined for sole-developer + AI-pair-programming workflows where the "CEO review" framing isn't directly applicable.

Related skills

• /brainstorm — earlier ideation stage
• /deep-interview — convergent requirement gating (predecessor)
• /verify — implementation verification (successor)
• /security-review — focused security audit (deeper than section 4 alone)