Multi-perspective plan review with VP Product, VP Engineering, and VP Design to identify blockers, anti-patterns, conflicts, and regressions.
Execution
When triggered, immediately:
- Read
agent-prompt.mdfrom this skill's directory:- User install:
~/.claude/skills/plan-review-skill/agent-prompt.md - Project install:
.claude/skills/plan-review-skill/agent-prompt.md
- User install:
- Spawn a single Task with
subagent_type: "general-purpose"passing the agent-prompt content as the prompt - Include in the prompt: the user's request, the plan file path or content, and current working directory
- Wait for the agent to complete
- Present the agent's summary to the user
Do NOT execute the review workflow in this session. The subagent handles VP agent spawning, issue consolidation, and user interaction (AskUserQuestion). Plan edits are returned as recommendations for the coordinator to apply — the subagent should not edit files directly, as background execution auto-denies Write/Edit tools.
Execution Context Requirements
This skill spawns a general-purpose subagent that coordinates VP review agents and consolidates findings.
Foreground execution required: Yes for interactive AskUserQuestion prompts. Background execution auto-denies clarifying questions.
Dispatcher tools (frontmatter): Read, Task, Bash (Bash is needed for the P-7 scanner fallback) Subagent tools: Read, Grep, Glob, Bash, Task, AskUserQuestion (Task to spawn the three VP agents; Bash to run the P-7 scanner). No Write/Edit.
Write/Edit delegation: The subagent returns approved plan modifications as text. The coordinator (main conversation) applies edits via its own Edit tool. This makes the skill resilient to all permission modes.
Reference: See Claude Code documentation on subagent tool permissions.
Bundled Scripts
scripts/scan-state-flip.sh is the scanner behind the P-7 State-Flip Assertion Audit. The review subagent runs it (via Bash) against the noun a change flips, e.g. bash <skill-dir>/scripts/scan-state-flip.sh <noun> [--ref <pre-flip-tree>]; run --help for scope options. It needs git and must run inside a git repository. If the subagent has no Bash access, the coordinator (which has Bash) runs the script and passes the output back.
The P-5 shared-state audit is always performed by hand; it does not depend on any extra tooling. If a concurrency-audit scanner (a separate skill) is installed, the reviewer must run it whenever P-5 triggers and reproduce its rows verbatim. Its output adds evidence and never replaces the manual checks.
Changelog
v2.1.0
- Rubric additions: P-5 shared-state audit with auto-populated matrix, P-6 observation/gate correctness, P-7 state-flip audit (
scripts/scan-state-flip.sh), P-8 negative pass criteria need both arms, P-9 wave sizing reflects the residue, ADR-worthiness check - P-5 manual checks are mandatory; a concurrency-audit scanner, if installed, is supplemental evidence only
v2.0.0
- Refactor: Thin dispatcher pattern — full logic extracted to agent-prompt.md
- Skill runs in isolated subagent context, significantly reducing post-compaction restoration overhead
See agent-prompt.md for the full changelog.