subagent-driven-development

name	subagent-driven-development
description	When the user approves a plan and wants to execute it. When the user says "execute the plan" or "let's go" or "start implementation". When a PLAN.md exists and is ready to be executed task-by-task.
tools	["read","exec","edit"]

Skill: Subagent-Driven Development

Purpose

Without task isolation, a single agent accumulates context across all tasks — leading to context bleed, hallucinations from earlier tasks, and difficult-to-debug failures. This skill prevents that by dispatching a fresh subagent per task with two-stage review before committing each result.

Auto-Trigger

• PLAN.md exists and user approves execution
• User says: "Execute the plan", "Let's go", "Start implementation"
• After prg plan generates a task list

Process

1. Dispatch Subagent

Isolate each task so earlier context cannot corrupt later work. Give the subagent: task description, relevant files, testing requirements — no knowledge of other tasks.

2. Subagent Executes

Let the subagent implement, write tests, and return code + test results before any review.

3. Stage 1 — Spec Compliance

Reject before checking quality; a non-compliant implementation wastes quality review time. ✓ Does it match task requirements? ✓ Are all files modified as specified? ✓ Do tests pass? ❌ If no → reject, provide feedback, retry

4. Stage 2 — Code Quality

✓ Follows project conventions? ✓ Handles edge cases? ✓ No obvious bugs? ✓ Code is readable? ❌ If no → request improvements

5. Commit and Report

Commit after each task so failures are isolated and rollback is cheap.

git add <files>
git commit -m "Task N: brief description"

✅ Task 1: Setup database schema (3min) 🔄 Task 2: In progress...

Stopping Conditions

• All tasks complete
• Critical issue found (escalate to user)
• Subagent fails 3 times (escalate)

Output

Each task produces: committed code, passing tests, and a progress entry in the session log.

Anti-Patterns

❌ Giving subagent entire codebase context ❌ Skipping Stage 1 review ❌ Continuing after failed tests ❌ Not committing after each task