BumpyClock/agent-templates / skills_archive/subagent-driven-development
subagent-driven-development
UnverifiedUse when executing a defined implementation plan with multiple mostly-independent tasks in the current session. Orchestrate implementer and reviewer subagents, parallelize safe work, keep architecture and integration in the orchestrator.
Subagent-Driven Development
Role: orchestrator. Split, dispatch, review, integrate, verify.
Orchestrator Owns
- Plan split.
- Interfaces, contracts, acceptance criteria.
- Architecture + product decisions.
- Dependency order.
- Integration.
- Final verification.
Core Rules
- Use
tasquefor durable tracking. - Split by ownership. Parallel agents must not edit same files or interfaces.
- Define interfaces + acceptance before dispatch.
- Give each subagent fresh, curated context: full task text, local context, owned files, contracts, constraints, tests. Do not make them infer from chat history.
- Keep decisions in orchestrator. Subagents implement, test, review, research.
- Require tests for behavior changes. Skip only for clearly mechanical work or explicit approval.
- Review every implementation with separate
reviewerpasses: spec compliance first, code quality second. - Loop till pass. If task stays stuck after repeated review/fix cycles, stop + surface blocker.
- Pick smallest capable implementer. Upgrade when signals show it is needed.
- Lint, format, test, live test instead of assuming success.
Implementer Selection
Default to developer-lite for mechanical tasks:
- Clear spec and acceptance criteria.
- Touches 1-2 files.
- No shared interfaces, schema, auth, security, concurrency, perf, or cross-module state.
- Existing pattern is obvious.
- Verification command is local and cheap.
Use developer for judgment tasks:
- Multi-file or cross-module work.
- Public API, schema, migration, state model, auth, security, concurrency, perf, or new dependency.
- Debugging, unclear patterns, large/tangled files, or architectural tradeoffs.
Upgrade rule: if developer-lite returns BLOCKED, NEEDS_CONTEXT, or correctness-related DONE_WITH_CONCERNS, add context, narrow the task, or redispatch to developer. Never retry unchanged.
Implementer Status
DONE: proceed to spec-compliance review.DONE_WITH_CONCERNS: read concerns. Resolve correctness/scope concerns before review; note observations and continue.NEEDS_CONTEXT: provide missing context and redispatch.BLOCKED: assess blocker. Add context, split task, upgrade implementer, or ask user if plan is wrong.
Flow
- Analyze
- Identify goal, success criteria, constraints, deps, shared surfaces, test needs.
- Break work into small tasks with clear ownership.
- Define contracts
- Lock public interfaces, data shapes, integration order, non-goals.
- Add enough detail for independent execution.
- Track
- Create or update
tsqtasks. - Model blockers explicitly.
- Mark active work
in_progress; close only after review + verification.
- Dispatch implementer
- Assign one owned task.
- Include task, context, owned files, contracts, constraints, tests, deliverables.
- Tell agent ask questions early. Avoid autonomous architecture changes.
- Select
developer-liteordeveloperusing Implementer Selection.
- Review
- Dispatch
reviewerwithReview mode: spec-compliance, requirements, changed files, diff context, test results. Do not trust implementer report. - After spec passes, dispatch
reviewerwithReview mode: code-quality. - If review fails, send targeted fixes back through an implementer.
- Integrate
- Merge approved work in dependency order.
- Re-run impacted tests after integration.
- Finish
- Run final end-to-end checks for plan.
- Dispatch final
reviewerwithReview mode: final-integrationover the full integrated diff. - Close
tsqitems with outcome notes. - Report blockers, residual risk, follow-up work.
Prompt Contract
- Full task statement. Paste it.
- Why task exists.
- Exact acceptance criteria.
- Owned files or module boundary.
- Interfaces or contracts to honor.
- Constraints + non-goals.
- Test expectations.
- Required deliverable format: status (
DONE|DONE_WITH_CONCERNS|BLOCKED|NEEDS_CONTEXT), summary, files changed, tests run, self-review findings, open issues.
References
- Red flags:
./references/guardrails.md - Example loop:
./references/example-workflow.md - Parallel split patterns:
./references/parallelization-examples.md
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 464k | Updated | Claude CodeCodex | Verified | Install |
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 464k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 318k | Updated | Claude CodeCodex | Unverified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 318k | Updated | Claude CodeCodex | Unverified | Install |