AgentSkillsCN

subagent-driven-development

在每项任务中，使用全新的子代理执行计划，并在任务间进行规格与代码质量的审查。

SKILL.md
--- frontmatter
name: subagent-driven-development
description: Execute plans using fresh subagents per task with spec and code quality reviews between tasks
user-invocable: false

Subagent-Driven Development

Overview

Execute implementation plans using fresh subagents for each task, with mandatory reviews between tasks. Three-phase cycle: implement → spec review → code quality review.

Core principle: Fresh context per task, skeptical reviews, fast iteration within session.

When to Use

  • •You have a completed implementation plan (from @writing-plans)
  • •Tasks are independent enough to be executed by fresh subagents
  • •You want fast iteration with quality gates in current session

The Three-Phase Cycle

For each task in the plan:

Phase 1: Implement

  1. •Dispatch fresh subagent with implementer-prompt.md template
  2. •Subagent reads plan, detects mode (TDD/Non-TDD), implements task
  3. •Subagent verifies work using plan's verification method
  4. •Subagent commits only after verification passes (using @committing)
  5. •Subagent returns completion report

Phase 2: Spec Review

  1. •Dispatch fresh subagent with spec-reviewer-prompt.md template
  2. •Subagent verifies implementation matches plan specs
  3. •Subagent re-runs the verification method from the plan
  4. •Returns PASS or FAIL with details
  5. •If FAIL: Loop back to Phase 1 with fixes

Phase 3: Code Quality Review

  1. •Only runs after Phase 2 PASS
  2. •Dispatch fresh subagent with code-quality-reviewer-prompt.md template
  3. •REQUIRED SKILLS:
    • •Use code-review-excellence for quality assessment
    • •Use code-simplifier to simplify and clean up code
  4. •Returns quality assessment with simplification suggestions
  5. •If issues found: Address before next task

Process Flow

code
┌─────────────────────────────────────────────────────────┐
│                    For Each Task                        │
├─────────────────────────────────────────────────────────┤
│                                                         │
│  ┌──────────────┐                                       │
│  │  Phase 1:    │                                       │
│  │  Implement   │──────────────────────┐               │
│  └──────────────┘                      │               │
│         │                              │               │
│         ▼                              │               │
│  ┌──────────────┐                      │               │
│  │  Phase 2:    │  FAIL                │               │
│  │  Spec Review │──────────────────────┘               │
│  └──────────────┘                                      │
│         │ PASS                                          │
│         ▼                                               │
│  ┌──────────────┐                                       │
│  │  Phase 3:    │                                       │
│  │  Code Review │                                       │
│  └──────────────┘                                       │
│         │                                               │
│         ▼                                               │
│     Next Task                                           │
└─────────────────────────────────────────────────────────┘

TodoWrite Integration

Create TodoWrite at start with structure:

code
Task 1: [Name from plan]
  - [ ] Phase 1: Implement
  - [ ] Phase 2: Spec Review
  - [ ] Phase 3: Code Quality Review
Task 2: [Name from plan]
  - [ ] Phase 1: Implement
  ...

Update after each phase completion.

Orchestrator Responsibilities

As the orchestrator (main agent), you:

  1. •Create TodoWrite from plan tasks
  2. •Dispatch subagents using Task tool with appropriate prompts
  3. •Review subagent reports between phases
  4. •Handle failures - dispatch fix attempts, escalate if stuck
  5. •Track progress - update TodoWrite after each phase
  6. •Report to user after each task cycle completes

Subagent Dispatch Pattern

code
Task tool call:
- subagent_type: "general-purpose"
- prompt: [Load and fill template from prompt file]
- Include: plan file path, task number, relevant context

Remember

  • •Fresh subagent = fresh context (no accumulated confusion)
  • •Spec review is skeptical - verify code, not claims
  • •REQUIRED SKILLS: Use code-review-excellence + code-simplifier for code quality reviews
  • •@committing for all commits
  • •Stop and escalate if spec review fails 3 times
  • •User can intervene between any phases