Subagent-Driven Development

Execute plans by dispatching fresh subagent per task, with two-stage review after each.

Core principle: Fresh subagent per task + two-stage review (spec then quality) = high quality, fast iteration.

Announce at start: "I'm using Subagent-Driven Development to execute this plan."

The Process

1. Read plan, extract all tasks with full text
2. Create TodoWrite with all tasks
3. Read prompt templates from this skill folder

For Each Task:
│
├─ Dispatch Implementer (./implementer-prompt.md)
│   ├─ If questions → answer, they continue
│   └─ When done → proceed to spec review
│
├─ Dispatch Spec Reviewer (./spec-reviewer-prompt.md)
│   ├─ If issues → implementer fixes → re-review
│   └─ When ✅ → proceed to quality review
│
├─ Dispatch Code Quality Reviewer (./code-quality-reviewer-prompt.md)
│   ├─ If Critical/Important → implementer fixes → re-review
│   └─ When approved → mark task complete
│
└─ Update TodoWrite, continue to next task

After All Tasks:
├─ Run final test suite
└─ Use finishing-branch skill

Prompt Templates

Read these files and use them to construct Task calls:

• ./implementer-prompt.md - Template for implementer subagent
• ./spec-reviewer-prompt.md - Template for spec compliance reviewer
• ./code-quality-reviewer-prompt.md - Template for code quality reviewer

When dispatching, use Task tool with subagent_type: general-purpose and fill in the template placeholders.

Example Flow

You: I'm using Subagent-Driven Development to execute this plan.

[Read plan file: docs/plans/feature-plan.md]
[Extract all 3 tasks with full text]
[Create TodoWrite with all tasks]
[Read ./implementer-prompt.md template]

Task 1: User validation

[Dispatch Task with implementer template filled in]

Implementer: "Should validation return 400 or 422?"

You: "Use 422 for validation errors."

Implementer:
  - Implemented validateUser
  - 6 tests, all passing
  - Self-review: Fixed missing email check
  - Committed

[Read ./spec-reviewer-prompt.md template]
[Dispatch Task with spec reviewer template]

Spec Reviewer: ✅ Spec compliant

[Read ./code-quality-reviewer-prompt.md template]
[Dispatch Task with quality reviewer template]

Code Quality Reviewer:
  Issues (Important): Magic regex should be constant

[Dispatch implementer to fix]

Implementer: Extracted EMAIL_REGEX constant, committed

[Re-dispatch quality reviewer]

Code Quality Reviewer: ✅ Approved

[Mark Task 1 complete]

Task 2: ...
[Continue pattern]

Key Behaviors

Read templates, don't hardcode prompts:

• Templates live in this skill folder
• Read them at execution time
• Fill in placeholders with task-specific content

Answer subagent questions:

• Implementers may ask before starting
• Provide clear, complete answers
• Don't rush them

Handle review loops:

• If reviewer finds issues → dispatch implementer to fix
• After fix → re-dispatch same reviewer
• Repeat until approved
• Never skip re-review

Track progress:

• Update TodoWrite as tasks complete
• One task in_progress at a time

Red Flags - NEVER

• Skip reviews (both required)
• Proceed with unfixed issues
• Dispatch parallel implementers
• Hardcode prompts instead of reading templates
• Make subagent read plan file (provide full text)
• Accept "close enough" on spec compliance
• Start quality review before spec passes
• Move to next task with open issues