Workflow Mastery Skill
Overview
This skill provides Claude with expert knowledge of the Claude Ultra workflow system, enabling automatic application of best practices for software development, task management, and context preservation.
When to Activate
Activate this skill when:
- Starting a new development session
- Working on complex multi-step tasks
- Managing git operations
- Performing code reviews
- Debugging issues
- Working with subagents
Workflow Protocol
Phase 1: UNDERSTAND
- Parse the request - what is actually being asked?
- Check constraints - tech stack, patterns, existing code
- Run
git status to understand current state
- Load relevant context
Phase 2: PLAN
- For complex tasks, use Plan Mode
- Use @explore agent to research unfamiliar code BEFORE modifying
- Break work into atomic tasks
- Get confirmation before major changes
Phase 3: EXECUTE
- Delegate to specialized subagents for distinct subtasks
- Use @task-worker agents in parallel for independent file operations
- Keep commits atomic - one logical change per commit
- Run tests/typecheck after each significant change
Phase 4: VERIFY
- Run the thing (tests, build, typecheck)
- Show output as evidence
- Review git diff before committing
- Never claim "done" without proof
Agent Selection Guide
| Task | Agent | When |
|---|
| Codebase search | @explore | Before modifying unfamiliar code |
| File modifications | @task-worker | Parallel independent changes |
| Code review | @code-reviewer | Before any PR |
| Bug investigation | @debugger | Root cause analysis |
| Security check | @security-auditor | Auth/data changes |
| Architecture | @architect | System design decisions |
Autonomy Modes
- Autonomous: Execute end-to-end with minimal check-ins
- Collaborative (default): Check in at major decision points
- Supervised: Explain each step before executing
Context Preservation
- Main agent handles orchestration
- Subagents handle detailed work
- This keeps main context clean for longer sessions
- Use memory-bank for persistent knowledge
Quality Gates
Before completing any task:
- Run the thing (tests, build, command)
- Show the output as evidence
- Verify it actually works