← ClaudeAtlas

keel-plan-reviewlisted

Use after keel-plan produces a plan for large or risky work — runs multi-lens automated review (business, engineering, plus design/DX/security when in scope) with adversarial verification, auto-deciding routine choices and escalating only genuine judgment calls. Stage 3 of the keel pipeline; hands off to keel-execute.
AWenSu/keel · ★ 0 · AI & Automation · score 75
Install: claude install-skill AWenSu/keel
# keel-plan-review — Multi-Lens Plan Review Rough plan in, reviewed plan out. Auto-decide only the Mechanical — findings with one defensible answer. Every genuine judgment call goes to the user as its own fully-detailed question, per the Step 5 rules. **Weight check:** this is the lean single-model version. For very large plans (>15 files, new product surface) where gstack is installed, prefer `/autoplan` — its dual-model consensus and mockup generation earn their cost there. ``` INPUT plan file with header (Goal, Spec Version, Global Constraints, Success Criteria) + tasks carrying Delivers / Files / Interfaces / Skills; the spec it derives from OUTPUT the same plan file, edited, with a REVIEW REPORT section ending in NO UNRESOLVED DECISIONS; deferrals written to TODOS.md ``` Missing INPUT → `BLOCKED: 缺 <field> → 退回 keel-plan`. Do not review a plan whose task fields are absent; the findings would have nothing to anchor to. ## Step 0: Premises — the ONE mandatory user question Before any review, confirm premises with the user in a single question: "This plan assumes <X, Y, Z>. Correct?" Wrong premises make every downstream finding worthless. This is the only question that must always be asked. ## Step 1: Detect scope → select lenses Always run: **CEO lens**, **Eng lens**. Conditionally add (2+ keyword hits in the plan): - **Design lens** — view/rendering/UI/component/screen vocabulary - **DX lens** — API/CLI/SDK/docs/MCP vocabulary (product i