subagent-driven-development

Subagent-Driven Development

Execute plans by dispatching a fresh subagent per task, with mandatory two-stage review to ensure spec compliance and code quality.

🎯 Load Full CORE Context

read ${PAI_DIR}/skills/CORE/SKILL.md

When to Activate This Skill

Decision Tree:

Do you have a plan?
├─ No → Use WritingPlans skill first
└─ Yes → Are tasks independent?
    ├─ No → Execute sequentially yourself
    └─ Yes → Can tasks be done in same session?
        ├─ No → User should open separate sessions (parallel execution)
        └─ Yes → USE THIS SKILL

Use this skill when:

• ✅ You have written plan (from WritingPlans or similar)
• ✅ Tasks are independent (Task B doesn't need Task A's output)
• ✅ All tasks can be done in same session (same codebase, same context)
• ✅ You want isolated context per task (prevents confusion)
• ✅ You want mandatory review checkpoints

Core Philosophy

Why Fresh Subagent Per Task?

Traditional approach: One agent does all tasks sequentially, carrying forward all context and decisions.

Problems:

• Context pollution: Earlier decisions affect later tasks inappropriately
• Assumption drift: Agent assumes things from earlier tasks apply to later ones
• Inconsistent quality: Agent gets tired/sloppy as session progresses
• No review: Agent reviews their own work (bias)

Subagent-Driven Approach: Fresh subagent for each task, two independent reviewers.

Benefits:

• ✅ Clean slate: Each implementer sees only their task spec
• ✅ Consistent quality: No fatigue effect across tasks
• ✅ Independent review: Reviewers haven't implemented, see with fresh eyes
• ✅ Natural TDD: Fresh context encourages test-first thinking
• ✅ Parallel potential: Tasks can be dispatched in parallel if needed

The Process

Step 1: Extract Tasks from Plan

Your job as orchestrator:

1. Read the plan document
2. Extract discrete tasks (each should be 2-5 minutes of work)
3. Verify tasks are independent
4. Create task list