| name | subagent-development |
| description | Use when executing implementation plans with independent tasks — dispatch fresh subagent per task with two-stage review (spec compliance, then code quality) for high quality, fast iteration. |
| metadata | {"version":"2.0.0","author":"erebus","tags":["software-development","subagents","implementation","automation"]} |
Subagent-Driven Development
Execute a plan by dispatching a fresh subagent per task, with two-stage review after each.
Why subagents: Fresh context per task = no context pollution. Precise instructions = focused success.
Each subagent gets exactly what it needs — not your full session history.
Core principle: Fresh subagent per task + two-stage review (spec then quality) = high quality, fast iteration.
When to Use vs. Executing Plans
Use subagent-development when:
- Tasks are mostly independent
- Subagent support is available
- You want isolated context per task
Use executing-plans instead when:
- No subagent support available
- Tasks are tightly coupled (must share state)
- Sequential execution is required
Architecture
Orchestrator (you)
├── Task 1 → Implementer subagent → Spec reviewer → Quality reviewer → ✓ merge
├── Task 2 → Implementer subagent → Spec reviewer → Quality reviewer → ✓ merge
└── Task 3 → Implementer subagent → Spec reviewer → Quality reviewer → ✓ merge
The Process
Setup
- Read the complete plan file
- Extract all tasks with their full text
- Note shared context (repo structure, tech stack, conventions)
- Create task list with status tracking
Per-Task Loop
For each task:
1. Dispatch implementer subagent
Give it:
- The COMPLETE task text from the plan (not a summary)
- Shared context (repo layout, test command, conventions)
- Clear "done" criteria
- Constraint: run tests, commit, self-review before returning
2. Handle questions
If the implementer asks questions: answer them, then re-dispatch.
3. Two-Stage Review
After implementer completes:
Stage 1 — Spec Compliance Review
Dispatch spec-reviewer subagent with:
- The original task spec
- The git diff of what was implemented
- Question: "Does the implementation match the spec? What's missing or wrong?"
If gaps found: send back to implementer to fix. Re-review.
Stage 2 — Code Quality Review
Dispatch code-quality-reviewer subagent with:
- The git diff
- Question: "Is this code clean, tested, and maintainable? Any issues?"
If issues found: send back to implementer to fix. Re-review.
4. Mark task complete
Only after both reviews pass.
After All Tasks
- Run the full test suite one final time
- Dispatch a final overall code reviewer for the complete implementation
- Use finishing-a-development-branch skill to complete the work
Two-Stage Review Detail
Stage 1: Spec Compliance
Check:
- Does the implementation address all requirements in the task?
- Are all specified files created/modified?
- Does the interface match what the spec described?
- Are there missing pieces?
Stage 2: Code Quality
Check:
- Is the code clean and readable?
- Are there tests? Do they test behavior (not just coverage)?
- Any obvious performance issues?
- Naming, duplication, complexity?
Rules
- Fresh subagent per task — never reuse a subagent across tasks
- Do NOT pass your full session history to subagents
- Both review stages are mandatory — never skip either
- Fix issues before marking a task complete
- If implementer is stuck after 2 attempts: escalate to you (the orchestrator) to unblock
Common Mistakes
Passing full session context — subagents need focused context, not everything you know.
Skipping review stages — "it looks fine" is not a review. Both stages required.
Moving on with open issues — fix Critical and Important issues before the next task.
Summarizing the task instead of quoting it — give the implementer the exact task text.