kienbui1995 Subagent-Driven Development
Execute plan by dispatching fresh subagent per task, with review after each.
Why subagents: Fresh context per task = focused work. Precisely crafted instructions = reliable results. Your context preserved for coordination.
Core principle: Fresh subagent per task + review = high quality, fast iteration.
When to Use
- Have an implementation plan with independent tasks
- Want to stay in current session
- Tasks can be worked on without shared state
vs. executing-plans: Same session, fresh context per task, review between tasks, faster iteration.
The Process
For Each Task:
Dispatch implementer subagent with:
- Task description from plan
- Relevant file paths and context
- Testing requirements
- Commit message format
Implementer works: implements, tests, commits
Dispatch reviewer agent (Haiku — fast, cheap):
- Review the diff
- Check spec compliance
- Check code quality
Handle review results:
- If approved → mark task complete, next task
- If issues → dispatch implementer to fix, re-review
Mark task complete in TodoWrite
Model Routing for Subagents
| Role | Agent | Model | Why |
|---|---|---|---|
| Implementation | (default subagent) | Sonnet | Good balance of quality/cost |
| Code review | reviewer | Haiku | Fast, cheap, read-only |
| Complex design questions | architect | Opus | Deep reasoning needed |
Subagent Instructions Template
When dispatching an implementer subagent, provide:
## Task
[Exact task from plan]
## Context
- Working directory: [path]
- Relevant files: [list]
- Dependencies: [what previous tasks created]
## Requirements
- Follow TDD: write test first, verify it fails, implement, verify it passes
- Commit after each step
- Run full test suite before reporting done
## Do NOT
- Modify files outside the task scope
- Skip tests
- Change architecture decisions from the plan
Completion
After all tasks:
- Use magic-powers:verification-before-completion
- Use magic-powers:finishing-a-development-branch
Integration
- magic-powers:writing-plans — creates the plan this skill executes
- magic-powers:requesting-code-review — review pattern used between tasks
- magic-powers:dispatching-parallel-agents — for truly independent tasks that can run concurrently