Subagent Driven Development

Subagent-Driven Development — Spawn, Review, Repeat

When to Use

Use when you have a written plan and need to execute it task by task using subagents.

Core Principle

One subagent per task + structured review after each = high quality, no context pollution.

Process

For Each Task in the Plan:

1. Spawn Subagent with Full Context

Provide the subagent with everything it needs upfront:

TASK: [Task title]

WHAT: [Exact description of what to build]

WHY: [Context — how this fits into the larger feature]

FILES:
- Create: [exact paths]
- Modify: [exact paths]

CODE:
[Complete code from the plan]

TESTING:
- Write failing test first (craftwork:test-driven-development)
- Run: [exact test command]
- Expected: PASS

CONSTRAINTS:
- [Project-specific constraints]
- [Security requirements]
- [Pattern compliance]

WHEN DONE:
1. Create MR/PR with descriptive title
2. Update task status to "review"
3. Report completion with verification evidence (craftwork:verification-before-completion)

2. Review the Output

After subagent reports completion, review the diff using craftwork:requesting-code-review:

• Matches plan specification (nothing more, nothing less)
• Tests exist and cover the feature
• No hardcoded secrets
• No breaking changes
• Code is clean and follows existing patterns

3. Approve or Request Fixes

If approved:

• Merge the MR/PR
• Mark task as done
• Move to next task

If issues found:

• Post specific feedback with file:line references
• Re-spawn subagent with fix instructions
• Do NOT skip the re-review after fixes

4. Move to Next Task

Only after current task is merged and verified. Never skip ahead.

Red Flags

• Never spawn multiple subagents on dependent tasks simultaneously
• Never skip review (even if subagent self-reviewed)
• Never proceed with unfixed issues
• Never merge without reading the diff
• Never let subagent read the plan file — provide full text instead