subagent-driven-development
Subagent-Driven Development
Execute plan by dispatching fresh subagent per task, with two-stage review after each: spec compliance review first, then code quality review.
Core principle: Fresh subagent per task + two-stage review (spec then quality) = high quality, fast iteration
When to Use
- Have an implementation plan with mostly independent tasks
- Want to stay in the current session (otherwise use
executing-plans) - See references/process-detail.md for decision tree and full process flow diagrams
The Process
- Load plan — Read
plan.md+manifest.json. Note task IDs and waves. Do NOT read briefing files into your context. - Dispatch per task — Send implementer a 2-3 sentence summary + briefing file path. Agent reads its own briefing from disk.
- Two-stage review — Spec compliance first, then code quality. Both must pass before marking complete.
- Wave boundaries — Re-read
manifest.jsonto recover state after context compaction.
Prompt Templates
./implementer-prompt.md—./spec-reviewer-prompt.md—./code-quality-reviewer-prompt.md
Red Flags
Never: skip reviews, proceed with unfixed issues, dispatch parallel implementers, paste full briefing text inline instead of pointing to briefing file, skip reading the manifest at wave boundaries, start code quality review before spec compliance passes.
- Dispatch prompts missing key sections (see skills-management/references/dispatch-prompt-template.md for the canonical 6-section structure)
If subagent asks questions: Answer before proceeding. If reviewer finds issues: Implementer fixes, re-review until approved.
Cleanup
After all tasks complete and final review passes, before calling git-workflow stack:
- Optionally write
.claude/plans/<plan-id>/summary.mdwith execution notes - Delete the plan directory:
rm -rf .claude/plans/<plan-id>/ - If deletion fails, warn but do not block
References
- references/workflow-example.md — full walkthrough
- references/advantages-and-costs.md — comparison vs manual/executing plans
Required skills: git-workflow, writing-plans, code-review, test-driven-development
More from hjewkes/agent-skills
skills-management
Use when creating, finding, installing, reviewing, or managing Claude Code skills — covers skill authoring, discovery, conventions, and lifecycle management
14github-pr
GitHub PR workflow — creating PRs, posting automated review comments, managing PR feedback cycles. Use when code is reviewed and ready for GitHub.
14buildkite
Buildkite CI/CD integration. Use when the user needs to check build status, trigger builds, read build logs, debug failures, manage pipelines, or any Buildkite workflow. Triggers include "buildkite", "build", "pipeline", "CI", "deploy", "build log", "build failed".
14sdk-verify
Verify Claude Agent SDK applications. Use after creating or modifying TypeScript or Python Agent SDK apps to check configuration and best practices.
11atlassian
Jira and Confluence integration for Atlassian Cloud. Use when the user needs to search issues, manage sprints, read Confluence docs, create/update pages, work with epics and backlogs, or any Atlassian workflow. Triggers include "jira", "confluence", "sprint", "backlog", "ticket", "issue", "epic", or references to Atlassian content.
11using-git-worktrees
Use when starting feature work that needs isolation from current workspace or before executing implementation plans - creates isolated git worktrees with smart directory selection and safety verification
8