A Claude Code agent workflow system that automates the software development lifecycle.
- Go to the local claude_setup repo
git checkout maingit pull
- Go to the New Project repo
rsync -av --exclude='.git' --exclude='README.md' <path to claude_setup> .
# Start development from a description
/dev Add rate limiting to the API
# Start from an existing issue
/dev 123
# Check progress
/dev status
# Resume after approving plan
/dev resume
# Cancel workflow
/dev abort- You describe what you want →
/dev Add user authentication - Architect agent plans → Creates issue with implementation checklist
- You review and approve → Comment "approved" or run
/dev resume - Agents execute autonomously:
- Coder → Implements via TDD (tests first)
- Reviewer → Security scan + code quality
- Tester → Runs tests, checks coverage
- Documenter → Updates docs and changelog
- SelfReflector → Captures learnings
- You review the PR → Merge when satisfied
Act as a Product Manager:
- Define what you want built
- Review and approve plans
- Review final PRs
The agents handle implementation details, testing, and documentation.
PLANNING → AWAITING_APPROVAL → CODING → REVIEWING → TESTING → DOCUMENTING → REFLECTING → DONE
- AWAITING_APPROVAL is the only human checkpoint
- Failed reviews/tests automatically retry (max 3 attempts)
- All state persisted in
.claude/tasks/for recovery
.claude/
├── CLAUDE.md # Entry point and conventions
├── settings.json # Permissions
├── agents/ # Agent definitions
│ ├── architect.md
│ ├── coder.md
│ ├── reviewer.md
│ ├── tester.md
│ ├── documenter.md
│ └── self-reflector.md
├── skills/ # Orchestration
│ ├── dev.md # Main workflow
│ └── git-workflow.md # Git conventions
├── learnings/ # Auto-updated knowledge
│ ├── patterns.md
│ ├── failures.md
│ └── preferences.md
└── tasks/ # Runtime state
- Branches:
<type>/<issue-number>-<description>(e.g.,feat/123-add-auth) - Commits: Conventional commits format
- Python: Ruff (lint/format), Pytest (testing), Bandit (security)
- Platforms: Auto-detects GitHub (
gh) or GitLab (glab)