← ClaudeAtlas

sdd-review-codelisted

SDD review when implementation must be audited against PRD, TechSpec, and tasks; does not correct findings.
dougcunha/TokenHound · ★ 3 · AI & Automation · score 79
Install: claude install-skill dougcunha/TokenHound
# Review SDD code 1. Require `prd.md`, `techspec.md`, and `tasks.md` under `tasks/prd-[slug]/`. Read PRD and TechSpec once per version; then manifest, tasks, and handoffs. Check every link, extra file, ID, state, and dependency. **Output:** every task has proven location and state; a missing source blocks review with the exact path. 2. Bound implementation by `--base` resolved to a commit, including relevant committed, staged, unstaged, and new files. Without a base, use handoffs and worktree; state the scope limitation. In re-review, include correction reports and handoffs without altering history. **Output:** reviewable set identified. An invalid base requests correction and never silently changes scope. 3. Build a matrix of every obligation: origin, implementation, test, state, and evidence. Reuse IDs; read each task's details according to its criteria without transcribing sources. Mark `conformant`, `non-conformant`, `pending`, or `not verifiable`. **Output:** no orphan obligation; incomplete tasks, broken links, and acceptance without evidence remain gaps. 4. Trace callers, effects, contracts, errors, and risks in changed paths; consult relevant rules/skills. Check commands and results; reuse execution proven for the same code/configuration/environment, running missing or invalidated checks. Run all TechSpec quality profile commands over the reviewable set, blocking and reservation, and subtract the Terrain baseline: debt that existed before the feature is n