autonomous-longtask-v2

Autonomous Long-Task Development

This skill optimizes Claude Code for long, autonomous development tasks—from multi-hour feature implementations to multi-session refactorings.

Core Principles

1. Loop Closing (Test-Driven)

Every code change must close a feedback loop:

1. Write test (define expected behavior)
2. Implement code
3. Run test → loop closed
4. Refactor if needed
5. Re-run test → confidence

Anti-pattern (Open Loop):

"Implement feature X"
→ Code written, no verification
→ Bugs discovered later

Best Practice (Closed Loop):

"Implement feature X with tests.
Run tests after implementation."
→ Immediate verification

2. Incremental Progress with Checkpoints

Never attempt to "one-shot" complex features:

❌ Implement entire feature at once
   → Context runs out mid-implementation
   → Next session inherits chaos

✅ Small, tested increments
   → Each increment works standalone
   → Clean handoff between sessions

Checkpoint patterns:

• Use /rewind or Esc Esc for rollbacks
• Commit after each working increment
• Document state in claude-progress.txt for session handoffs

3. Context Management

Claude has 200K tokens, but:

• Subagents have isolated context windows
• Long sessions fragment context
• Use claude-progress.txt for session handoffs
• /compact <focus> for manual compaction

---

Task System

Use TaskCreate/TaskUpdate for persistent progress tracking (replaces deprecated TodoWrite).

Quick Start

1. TaskCreate: Create task with subject, description, activeForm
2. TaskUpdate: Set dependencies with addBlockedBy
3. TaskUpdate: Mark in_progress when starting
4. TaskUpdate: Mark completed when done
5. TaskList: Find next available work

Task Workflow

pending → in_progress → completed
         ↑
    (start work)

Tasks persist across sessions and auto-unblock when dependencies complete.

For detailed patterns: See references/task-patterns.md