feature

name

feature

description

Build ONE medium-size feature autonomously and hands-off — spec (spec-writer → spec-reviewer) → auto-approve → plan → implement (TDD) → verify gate → adversarial panel → e2e, with opinionated medium defaults and NO launch questions. Use when you've described a single feature and want it built end-to-end while you're away, without being pinged. The mid-tier between interactive create-spec+plan and the big build loop. A thin recipe over the author-specs + build-feature workflows — no new engine, no new agents.

feature — build one medium feature, hands-off

The mid-tier entry point. build is the big autonomous loop (a whole roadmap / a todo list, full thoroughness); create-spec + plan mode is interactive (you review every gate). feature is the rung between: one feature, described in a line, built end-to-end without touching you — spec → approve → plan → implement → verify → panel → e2e — with medium defaults baked in so there are no launch questions.

It is a thin playbook the main session runs, not a new machine: it chains the existing author-specs and build-feature Workflows. (Like build, the orchestrator stays the main session — a hands-off Workflow can't answer a blocking clarification.) For the machinery details, read .claude/skills/build/SKILL.md; this skill only sets the single-feature, medium-default wiring.

The pipeline (given /feature "<one-line brief>")

1. Already Approved? If an Approved docs/<feature>_system_design.md already exists for this feature → skip to step 4 (don't re-author).

2. Spec. Launch the author-specs Workflow with items:[{feature, brief}] (spec-writer drafts + self-reviews → spec-reviewer independent gate). Single-item run of the existing front.

3. Approve (hands-off; auto-approve + veto post-hoc):

   • ready-to-approve → flip Status: Approved + commit the spec (docs(specs): …).
   • ready-with-fixes (warnings only, no blocker) → auto-approve + commit; surface the warnings in the final report (assume-and-flag — you decide post-hoc if the warning mattered).
   • not-ready (a blocker / a live [blocking] open question) → PARK: commit the Draft spec, stop, and lead the final report with the clarification batch. Do not guess a blocking design decision while the author is away — a [blocking] means the plan can't be assembled without the answer. (No ping; the batch is waiting on return.)

4. Build. Read the approved spec's ## Success Criteria + ## Tests to assemble scs:[{id,summary}], then launch the build-feature Workflow with medium defaults:

   { feature, spec, scs:[{id,summary}], verifyCommand:"<the repo's strict verify gate>",
     maxIters:3, e2eBackstop:3 }
   

   fullCycle omitted → lean cycle (per-SC evaluator OFF, high-signal skeptic ON); planReview left to its complexity-gated default (ON for ≥3 SCs); sessionUrl passed if available (commit trailer); invariants/lenses/migrationHeadsCommand/liveTestsCommand passed if the repo defines them. No launch questions — the point is fire-and-forget. Iteration budget is deliberately medium (3, not converge/8): bounded thoroughness, not exhaustive.

5. Report (only at the end — never mid-run): what was built, the SC→test map, the final gate status, e2e evidence (a real-render screenshot per user-facing SC), any assumptions/warnings it flagged, and converged/done.

Defaults & scope (opinionated on purpose)

• Medium thoroughness: lean cycle + iteration budget 3. Want maximum scrutiny or a bigger budget → use build with {fullCycle:true} / converge. Want to babysit → create-spec + plan mode.
• One feature only. No scope creep into siblings. A brief that's really several features → that's a build job (roadmap + front), not feature.
• Commits autonomously (needs the same git allow-list as build: git add/commit/switch/merge; the deny-list stays). Never pushes.

Gates kept

• Spec readiness — the spec-reviewer gate must clear before auto-approve (a not-ready parks).
• Every increment — the verify gate green (the build-feature gate) + the whole-spec adversarial panel + e2e until proven.
• Every user-facing SC — a ui-verify real-render screenshot opened and ruled clean (docs/agent_rules/testing.md §5), else not done.

Reference

• The machinery: .claude/skills/build/SKILL.md + .claude/workflows/{author-specs,build-feature}.workflow.js.
• Which entry point for which task: docs/index.md ("What you're doing → what to read").