Subagent Dev

Execute a plan using subagent-driven development for: $ARGUMENTS

Iron Law

> The controller coordinates. Subagents implement. Never mix roles.

The controller agent (you) MUST NOT write production code directly. You dispatch subagents for each task, review their work, and coordinate the overall plan.

Architecture

Controller (you)
  ├── Read the plan
  ├── For each task:
  │   ├── Dispatch Implementer subagent (fresh context)
  │   ├── Dispatch Spec Reviewer subagent
  │   ├── Dispatch Quality Reviewer subagent
  │   └── Handle status (DONE / CONCERNS / BLOCKED)
  └── Verify all tests pass at the end

Steps

1. Load the Plan

Read task_plan.md or the user's task list. Extract all tasks with:

• Task description
• Files involved
• Dependencies (what must complete first)
• Acceptance criteria

2. Execute Each Task

For each task (respecting dependency order):

A. Dispatch Implementer

Launch a subagent with Agent tool:

Prompt: "Implement [task description].
Files to modify: [list]
Pattern to follow: [reference file]
Acceptance criteria: [criteria]
Run tests after implementation."

Use model: "sonnet" for mechanical tasks, model: "opus" for complex architecture.

B. Check Implementer Status

The implementer should report one of:

| Status | Meaning | Action | |--------|---------|--------| | DONE | Task complete, tests pass | Proceed to review | | DONEWITHCONCERNS | Complete but has questions | Review concerns, decide if acceptable | | NEEDS_CONTEXT | Missing information to proceed | Provide context, re-dispatch | | BLOCKED | Cannot complete (dependency, bug) | Log blocker, skip to next task |

C. Two-Stage Review

Stage 1 — Spec Compliance (must pass first):

• Does the implementation match the task description?
• Are all acceptance criteria met?
• Were the correct files modified?

Stage 2 — Code Quality (only after Stage 1 passes):

• Does it follow existing codebase patterns?
• Are there security issues?
• Is the code clean and maintainable?

If either review fails, send feedback to a new implementer subagent for fixes.

3. Verify Everything

After all tasks are done:

# Run full test suite
[test command]

# Check for uncommitted changes
git status

4. Report

## Subagent Execution Report

### Tasks Completed: X/Y
| Task | Status | Implementer | Review |
|------|--------|-------------|--------|
| [task] | DONE | Pass | Pass |

### Files Changed
- [file list]

### Concerns Raised
- [any DONE_WITH_CONCERNS items]

### Blocked Items
- [any BLOCKED items with reasons]

Anti-Rationalization

| Excuse | Rebuttal | |--------|----------| | "This task is too small for a subagent" | Small tasks are the BEST use of subagents — they get clean context | | "I'll just make this quick fix myself" | You are the controller. Controllers don't write code. | | "The review is overkill for this change" | Every change gets reviewed. No exceptions. | | "I'll review everything at the end" | Review after EACH task. Catching issues early is 10x cheaper. | | "The tests were passing before my changes" | Verify. Don't assume. Run them. |

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: huzaifa525
• Source: huzaifa525/claude-code-optimizer
• License: MIT
• Homepage: https://www.npmjs.com/package/claude-code-optimizer

Install and usage instructions live in the source repository linked above.