Subagent-Driven Development

Overview

Execute implementation plans by dispatching independent subagents for each task, with a two-stage review process.

Core principle: Fresh subagent per task + two-stage review (spec then quality) ensures quality and speed.

The Process

Step 1: Extract Tasks

Load plan file and extract all tasks into TodoWrite tracker.

Step 2: For Each Task

1. Implementation: Dispatch an implementer subagent who:

   • Codes the solution
   • Tests the implementation
   • Self-reviews

2. Spec Review: A separate subagent verifies:

   • Code matches requirements
   • All acceptance criteria met
   • No over/under-building

3. Quality Review: Another subagent assesses:

   • Code quality
   • Best practices
   • Maintainability

4. Iterations: If either review identifies gaps:

   • Implementer fixes issues
   • Both reviewers re-check

Key Principles

• Fresh subagent per task - prevents context pollution
• Two-stage review - spec compliance before code quality
• Same session - eliminates context switching
• Surface questions early - before implementation begins

Critical Safeguards

• Never skip reviews
• Never accept "close enough" on spec compliance
• Spec compliance must be verified BEFORE code quality review
• Order matters: spec then quality

Why This Approach

Rather than:

• Manual execution (slow, error-prone)
• Parallel sessions (context loss)

This provides:

• Parallelization via independent task execution
• Fresh context per task
• Systematic verification at each stage

Common Mistakes

• Skipping spec review
• Reviewing quality before spec compliance
• Using same subagent for implementation and review
• Batching multiple tasks without per-task review

Red Flags - STOP

• "This task is simple, skip review"
• "Quality review first, spec later"
• Same subagent does implementation and review
• "Close enough" on requirements