Subagent Driven Development

This skill does NOT contain agent prompts. It tells the orchestrator the dispatch pattern. The agents define their own behavior and use their own skills.

The Process

Setup Phase

1. Read plan file once
2. Extract all tasks with full text and context
3. Create TodoWrite tracking all tasks

Per-Task Cycle

For each task:

┌─────────────────────────────────────────────────────────────┐
│  1. IMPLEMENTER (use .claude/developer.md)                  │
│     ├─ Dispatch with full task context                      │
│     ├─ Answer any clarifying questions                      │
│     ├─ Developer: implement → test → commit → self-review   │
│     └─ Returns: implementation report with verification     │
│                                                             │
│  2. SPEC REVIEWER (use .claude/reviewer.md Stage 1)         │
│     ├─ "Did we build what was specified?"                   │
│     ├─ Compare implementation against spec line-by-line     │
│     ├─ If issues → implementer fixes → re-review            │
│     └─ Returns: ✅ Spec compliant OR ❌ Issues (with refs)   │
│                                                             │
│  3. CODE QUALITY REVIEWER (use .claude/reviewer.md Stage 2) │
│     ├─ "Is it well-built?"                                  │
│     ├─ Review for quality, performance, maintainability     │
│     ├─ If issues → implementer fixes → re-review            │
│     └─ Returns: ✅ Approved OR ❌ Issues (with refs)         │
│                                                             │
│  4. Mark task complete in TodoWrite                         │
└─────────────────────────────────────────────────────────────┘

Dispatching Agents

For Implementer:

Task(
  subagent_type="general-purpose",
  description="Implement [task name]",
  prompt="""
Use @test-driven-development and @verification-before-completion skills.

TASK CONTEXT:
- Task: [task name and description from plan]
- Files: [relevant files]
- Dependencies: [what this depends on]

IMPLEMENT THIS TASK with TDD: write failing test, implement, verify pass.
"""
)

For Spec Review:

Task(
  subagent_type="review",
  description="Spec review [task name]",
  prompt="""
Perform STAGE 1 (Spec Compliance) review only.

SPECIFICATION:
[Original task spec from plan]

IMPLEMENTATION:
[Implementer's report]

Check: Did we build what was specified? No more, no less.
"""
)

For Quality Review:

Task(
  subagent_type="review",
  description="Quality review [task name]",
  prompt="""
Perform STAGE 2 (Code Quality) review.

IMPLEMENTATION:
[Files changed, code to review]

PREVIOUS REVIEW:
[Spec review passed]

Check: Is it well-built? Readability, error handling, tests.
"""
)

Completion Phase

After all tasks:

1. Dispatch final code reviewer for entire implementation
2. Run full test suite
3. Mark implementation complete

Two-Stage Review: Why This Order?

WRONG ORDER:
  Code quality first → "This is well-written"
  Spec review later  → "But it's not what we asked for"
  Result: Wasted quality review on wrong code

RIGHT ORDER:
  Spec review first  → "This matches the specification"
  Code quality after → "And it's well-written"
  Result: Quality review only on correct code

Spec reviewer catches:

• "You added unrequested parameters" (YAGNI)
• "You skipped the edge case in the spec"
• "Output format doesn't match requirements"

Code quality reviewer catches:

• "This loop should be optimized"
• "Error handling is incomplete"
• "This pattern doesn't match codebase conventions"

Example Workflow

Initial Setup

You: Using Subagent-Driven Development for this plan.
[Read plan, extract 5 tasks, create TodoWrite]

Task Execution

Task 1: Add input validation to user service

[Dispatch general-purpose agent with task context]

Implementer: "Should validation raise exception or return error object?"
You: "Raise ValueError for invalid input."

Implementer: [Implements with TDD, tests 3/3 pass, self-reviews, commits]

[Dispatch review agent for Stage 1 - Spec Compliance]
Reviewer: ✅ Spec compliant
  - Input validation implemented
  - ValueError raised for invalid input
  - Tests cover edge cases

[Dispatch review agent for Stage 2 - Code Quality]
Reviewer: ✅ Approved
  - Clean implementation
  - Good error messages
  - No issues found

[Mark Task 1 complete]

With Review Iterations

Task 2: Add rate limiting to API endpoints

[Implementer implements, tests 5/5 pass, commits]

[Dispatch review agent Stage 1]
Reviewer: ❌ Issues:
  - Missing: per-user rate limiting
  - Extra: added caching parameter (not in spec)

[Implementer fixes - removes caching, adds per-user limits]
[Reviewer re-reviews Stage 1]
Reviewer: ✅ Compliant now

[Dispatch review agent Stage 2]
Reviewer: ❌ Issue: Magic number (100) should be constant

[Implementer extracts DEFAULT_RATE_LIMIT constant]
[Reviewer re-reviews Stage 2]
Reviewer: ✅ Approved

[Mark Task 2 complete]

Red Flags

Never:

• Skip spec review (even if "obviously correct")
• Skip code quality review
• Proceed with unfixed issues
• Let implementer self-review replace actual review
• Start code quality review before spec passes
• Move to next task with open review issues

If subagent asks questions:

• Answer clearly and completely
• Provide additional context if needed
• Don't rush into implementation

If reviewer finds issues:

• Same implementer fixes them
• Reviewer reviews again
• Repeat until approved

Integration with Other Skills

Implementer agent uses:

• @test-driven-development — For implementing each task
• @systematic-debugging — If tests fail during implementation
• @verification-before-completion — Before claiming task done

Review agent uses:

• @code-reviewer — For domain-specific quality checks (if available)

When receiving review feedback:

• @receiving-code-review — For responding to findings