Feature Work

feature-work

Second stage of the 4-skill feature workflow. Consumes a Plan approved by feature-research and produces working code, guided by the Plan's ## Task Order and ## Change Plan sections.

$ARGUMENTS must be the path to an existing Plan file. If missing, resolve via AskUserQuestion.

> Language note: This skill is authored in English per .claude/rules/skill-authoring.md. User-facing output follows the project's CLAUDE.md language preference.

Invariants

• Plan body is read by path, never embedded into agent prompts. Each dispatched agent opens the Plan via Read itself.
• User questions use AskUserQuestion.
• Parallel dispatch only when the target files of two tasks do not overlap (checked against the Plan's ## Change Plan table).
• Every delegated agent prompt ends with the fixed return schema from feature-research/references/plan-schema.md (§ Agent Return Schema).

References

• [references/task-breakdown.md](references/task-breakdown.md) — Task decomposition and parallel dispatch examples.
• ../feature-research/references/plan-schema.md — Plan file contract and agent return schema.

---

Workflow

1. Resolve the Plan

1. Take the Plan path from $ARGUMENTS. If empty, list files under .claude/plans/ and ask the user which to use via AskUserQuestion. If none exist, stop and recommend running /feature-research first.
2. Read the Plan's frontmatter. If status.research is pending, ask (AskUserQuestion) whether to (a) abort, (b) continue anyway, or (c) run /feature-research first.

2. Decompose into tasks

1. Parse the Plan's ## Task Order section. For each entry:

• TaskCreate.subject = the entry's title
• TaskCreate.description = the entry's **Detail** line

1. Establish dependencies from the **Depends on** lines using TaskUpdate.addBlockedBy.

3. Execute tasks

Loop: pick ready tasks (no unfinished blockers), dispatch them, advance.

• Single task ready: TaskUpdate → in_progress, delegate to the implementer agent (subagent_type: implementer) with:
• A pointer to the Plan path and the task's title/detail.
• The specific files this task touches (from ## Change Plan).
• The reuse hints relevant to this task (from ## Reuse).
• The scoped test command for the touched paths (if available).
• The fixed return schema (so only a summary comes back).
• Multiple ready tasks with disjoint file sets: dispatch them in a single Agent block as parallel implementer instances. See references/task-breakdown.md for shape.
• File overlap detected: serialize; do not dispatch in parallel.

The implementer agent (see agents/implementer.md) loads the Plan by path, performs the edits, runs scoped tests internally with up to 2 self-heal retries, and returns only the fixed return schema.

After each task's agent returns status: ok, call TaskUpdate → completed.

4. Targeted tests after each task

Run the scoped test command for the touched paths after each task (e.g., vitest , pytest , go test ). Long test output should be captured by the agent that owns the task and only summarized to main. Detect the command from package.json / Makefile / build.gradle / pyproject.toml / CLAUDE.md.

5. Full test sweep + simplify

1. Once all tasks are completed, run the project's full test command once to catch regressions.
2. Invoke /simplify to let the review pass refine reuse/quality/efficiency on the changed code.
3. If /simplify makes edits, re-run the test sweep.

6. Update Plan status

Edit the Plan file's frontmatter: set status.work: done. Do not touch other fields.

7. Mini-report (outline style)

Emit a final outline-form report as model text, in the project's preferred language. Labels below are reference-only.

──────────────────────────────────────────
  ✓ feature-work done   ****
──────────────────────────────────────────

**▸ Summary**
  - **Tasks executed** — ****
  - **Parallel dispatches** — ****
  - **simplify** — ✓ applied

**▸ Artifacts**
  - `` — ****
  - …

**▸ Metrics**
  - Tests ****
  - Files changed ****

**▸ Next steps**
  - **Run `/feature-test `** to add and run tests
──────────────────────────────────────────

Outline rules: no complete sentences; bold nouns/identifiers/numerics/status; symbols only from ✓ ⚠ ✗ ▸; no full path dumps. Never emit via Bash — model text only.

Constraints

• SKILL.md body length ≤ 200 lines.
• Do not delete or re-order Plan sections; touch only status.work.
• Do not start feature-test or feature-review from here; the wrapper orchestrates transitions.
• Two retries maximum for agent self-healing on a single task; on a third failure surface to the user.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Tianea2160
• Source: Tianea2160/claude-skills
• License: MIT

Install and usage instructions live in the source repository linked above.