← ClaudeAtlas

implementlisted

Run the red-green TDD loop on the current slice of a feature spec. Use to start or resume building a spec, or when the user says continue working on a feature. Stops at the slice boundary for human review. Phase 3 of the build pipeline.
Ashutosh6393/Scaffold-AI · ★ 0 · Testing & QA · score 57
Install: claude install-skill Ashutosh6393/Scaffold-AI
# implement — Phase 3 Build the **current in-progress slice** of a spec, and nothing else. This is the only phase that runs without a human in the inner loop, so the discipline is strict and partly machine-enforced (`guard-paths.sh`, `check-test-count.sh`). --- ## Step 1 — locate the spec **Named** (`implement notifications`) → use it. **Not named** → list the state and ask: ```bash for d in specs/*/; do [ "$(basename "$d")" = "_templates" ] && continue printf '%-32s %s\n' "$(basename "$d")" "$(grep -m1 'Status:' "$d/implementation.md")" done ``` Show in-progress specs first, then not-started, and ask which to work on. Don't guess. **If no spec exists** → stop. Tell the user to run `to-spec`. Do not start implementing from an ADR directly; the slice plan is what makes the loop bounded. --- ## Step 2 — orient 1. Read `specs/{feature}/CLAUDE.md`, then `design.md` (slice plan), then `implementation.md`. 2. Identify the **current slice**, its **blast radius**, and its **acceptance criteria**. 3. Confirm you are on `feat/{feature-name}`. 4. **If any task is `blocked`, stop and report it.** Do not start other tasks — they are dependency-ordered, and working around a block means building on something known to be broken. Record the starting test count so the append-only rule is checkable: ```bash bash .claude/hooks/check-test-count.sh --baseline ``` --- ## Step 3 — the loop One task at a time, in dependency order. One cycle at a time. ### RED — delegate