| name | subagent-driven-development |
| description | Use when executing implementation plans with independent tasks - dispatch fresh subagent per task with two-stage review |
Subagent-Driven Development
Execute plan by dispatching fresh subagent per task, with two-stage review after each: spec compliance first, then code quality.
Core Principle
Fresh subagent per task + two-stage review (spec then quality) = high quality, fast iteration
The Process
- Read plan, extract all tasks with full text, create todo list
- For each task:
a. Dispatch implementer subagent with full task text + context
b. Handle implementer status (DONE / DONE_WITH_CONCERNS / NEEDS_CONTEXT / BLOCKED)
c. Dispatch spec reviewer - does code match spec?
d. If issues -> implementer fixes -> re-review
e. Dispatch code quality reviewer - is implementation well-built?
f. If issues -> implementer fixes -> re-review
g. Mark task complete
- After all tasks: dispatch final code reviewer for entire implementation
Implementer Status Handling
- DONE: Proceed to spec review
- DONE_WITH_CONCERNS: Read concerns, address if correctness-related, note observations
- NEEDS_CONTEXT: Provide missing context, re-dispatch
- BLOCKED: Assess - context problem? Re-dispatch. Too large? Break up. Plan wrong? Escalate to human.
Red Flags - Never
- Skip reviews (spec compliance OR code quality)
- Proceed with unfixed issues
- Dispatch multiple implementation subagents in parallel (conflicts)
- Let implementer self-review replace actual review
- Start code quality review before spec compliance passes
- Move to next task while either review has open issues
Model Selection
- Mechanical tasks (1-2 files, clear spec) -> fast, cheap model
- Integration/judgment tasks (multi-file, debugging) -> standard model
- Architecture/design/review tasks -> most capable model
