Feature

---
description: "Build feature autonomously: research -> plan -> approve -> work -> review -> compound"
category: workflow
version: 1.0.0
argument-hint: "[feature description]"
---

# Build Feature: $ARGUMENTS

Follow the build-feature skill in `.claude/skills/build-feature/SKILL.md` exactly.

## Variables

- `$ARGUMENTS`: Feature description to build

## Phase 0: Resume Detection

**ALWAYS run this first.** Read existing Atlas state files:

```bash
cat .claude/atlas/current-feature.txt 2>/dev/null || echo "(none)"
cat .claude/atlas/STATE.md 2>/dev/null || echo "(none)"
cat .claude/atlas/progress.txt 2>/dev/null || echo "(none)"

If resuming same feature: Present options (continue / re-plan / start fresh) and wait for user choice. If fresh start: Archive any existing progress, then proceed to Phase 1.

Phase 1: Research

Spawn parallel Explore agents to understand the codebase:

1. Find similar patterns in src/
2. Check git history for related changes
3. Review CLAUDE.md and any AGENTS.md files

Synthesize findings into brief context.

Phase 2: Plan

1. Setup Atlas files:
   • Save feature name to .claude/atlas/current-feature.txt
   • Initialize .claude/atlas/STATE.md with position and context
   • Reset .claude/atlas/progress.txt
2. Create TodoWrite task list with detailed acceptance criteria
3. Document any architectural decisions needed

Phase 3: Present for Approval

Show the complete plan including:

• Research summary
• Task list with dependencies
• Acceptance criteria
• Risks and mitigations

WAIT for user to say "approved" before continuing.

Phase 4: Work (after approval only)

Execute via Atlas:

• Pick ready tasks from TodoWrite
• Decide: inline (simple) vs sub-agent (complex)
• Implement following project patterns
• Run echo "Validation passed - shell scripts and markdown" after each task
• Update .claude/atlas/STATE.md with position
• Append learnings to .claude/atlas/progress.txt
• Commit after each task
• Repeat until all tasks complete

Phase 5: Review

1. Run code-reviewer agent on all changes
2. For UI changes, use Playwright to verify functionality (if available, skip if not installed)

Phase 6: Compound

Ask: What did we learn?

• Project-wide patterns -> CLAUDE.md
• Folder-specific gotchas -> src/[folder]/AGENTS.md
• Decisions made -> CLAUDE.md Decisions section

Phase 7: Final Summary

Report:

• What was built
• Files changed
• Commits made
• Learnings captured
• Ready for PR

Complete

When this command completes:

• Feature fully implemented and validated
• All tasks complete in TodoWrite
• Code reviewed and functional verification passed (if applicable)
• Learnings captured in CLAUDE.md / AGENTS.md
• Commits made for each task
• Ready for: PR creation or user review