菜单
Manage autonomous task completion workflow. Use this at the START of any task that requires building, creating, fixing, or implementing something.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Convert files between formats (PDF to text, Markdown to HTML, JSON to CSV, CSV to JSON). Use when asked to convert, transform, or export files to different formats.
Persistent memory with semantic search - store and recall information across sessions. Use when asked to remember something, recall previous discussions, or manage long-term memory. Also use proactively to store important discoveries and compact context.
Install software packages at runtime in the Docker container. Use when a required tool or library is missing, or when asked to install packages.
Query SQLite and PostgreSQL databases. Use when asked to run SQL queries, analyze database data, or explore database tables.
MANDATORY final step. Validates task completion through testing and QA before allowing stop.
Track tasks for the current work session. Use this to break down work into steps and track progress. ALWAYS use this when starting a new task - add items first, then mark them done as you complete them.
| name | workflow |
| description | Manage autonomous task completion workflow. Use this at the START of any task that requires building, creating, fixing, or implementing something. |
ALWAYS use /workflow start when the user request is a TASK, not a chat.
| Request Type | Example | Action |
|---|---|---|
| Chat | "What is React?" | Answer directly, NO workflow |
| Chat | "Explain this code" | Answer directly, NO workflow |
| Task | "Build a todo app" | /workflow start FIRST |
| Task | "Fix the login bug" | /workflow start FIRST |
| Task | "Add dark mode" | /workflow start FIRST |
| Task | "Create an API endpoint" | /workflow start FIRST |
Rule: If you will write/edit code or create files → USE WORKFLOW.
┌─────────────────────────────────────────────────────────────────┐
│ USER REQUEST │
└─────────────────────────────────────────────────────────────────┘
│
▼
┌─────────────────┐
│ Chat or Task? │
└─────────────────┘
│ │
Chat ───┘ └─── Task
│ │
▼ ▼
┌──────────┐ ┌─────────────────────┐
│ Answer │ │ /workflow start │
│ directly │ │ "task description" │
└──────────┘ └─────────────────────┘
│
┌───────────────────┘
▼
┌─────────────────────┐
│ PHASE 1: PLAN │
│ │
│ Think through: │
│ • What to build │
│ • Key components │
│ • Dependencies │
│ • Testing approach │
└─────────────────────┘
│
▼
┌─────────────────────┐
│ PHASE 2: TASKS │
│ │
│ /tasks add "..." │
│ /tasks add "..." │
│ /tasks add "..." │
│ │
│ Break plan into │
│ trackable steps │
└─────────────────────┘
│
▼
┌─────────────────────┐
│ PHASE 3: WORK │◄────────────────────┐
│ │ │
│ /tasks working N │ │
│ • Implement │ │
│ • Use subagents │ │
│ if stuck │ │
│ /tasks done N │ │
└─────────────────────┘ │
│ │
▼ │
┌─────────────────────┐ │
│ PHASE 4: TEST │ │
│ │ │
│ • Run tests │ │
│ • Fix failures │ │
│ • If UI project: │ │
│ /vision verify │ │
└─────────────────────┘ │
│ │
▼ │
┌─────────────────────┐ │
│ More tasks? │───── Yes ──────────┘
└─────────────────────┘
│
No
│
▼
┌─────────────────────┐
│ PHASE 5: QA │
│ │
│ Call qa-agent: │
│ • Verify coverage │
│ • Check all tested │
│ • Identify gaps │
└─────────────────────┘
│
┌───────┴───────┐
▼ ▼
QA VERIFIED GAPS FOUND
│ │
│ ▼
│ ┌───────────┐
│ │ Add tasks │
│ │ for gaps │──────────────────┐
│ └───────────┘ │
│ │
▼ │
┌─────────────────────┐ │
│ PHASE 6: SUPERVISOR │ │
│ │ │
│ /supervisor │ │
│ • Checks tasks │ │
│ • Checks tests │ │
│ • Checks QA │ │
│ • Makes decision │ │
└─────────────────────┘ │
│ │
┌──────────┼──────────┐ │
▼ ▼ ▼ │
ALL COMPLETE CONTINUE RE-PLAN │
│ │ │ │
│ └──────────┼────────────────────┘
│ │
▼ ▼
┌──────┐ ┌──────────┐
│ STOP │ │ Back to │
│ │ │ PLAN │
└──────┘ └──────────┘
# Start workflow for a new task
~/.claude/skills/workflow/workflow.sh start "task description"
# Record a checkpoint (auto-called by other skills)
~/.claude/skills/workflow/workflow.sh checkpoint plan
~/.claude/skills/workflow/workflow.sh checkpoint tasks
~/.claude/skills/workflow/workflow.sh checkpoint test
~/.claude/skills/workflow/workflow.sh checkpoint vision
# Check current workflow state
~/.claude/skills/workflow/workflow.sh status
# Reset workflow (start over)
~/.claude/skills/workflow/workflow.sh reset
Before writing any code, think through:
What exactly needs to be built?
What components are needed?
What's the testing approach?
Any research needed?
web-researcher subagent if neededBreak your plan into trackable tasks:
~/.claude/skills/tasks/tasks.sh add "Setup: Create project structure"
~/.claude/skills/tasks/tasks.sh add "Core: Implement main logic"
~/.claude/skills/tasks/tasks.sh add "UI: Create user interface"
~/.claude/skills/tasks/tasks.sh add "Test: Write and run tests"
~/.claude/skills/tasks/tasks.sh add "Verify: Visual verification"
Good tasks are:
For each task:
# Mark as in progress
~/.claude/skills/tasks/tasks.sh working 1
# Do the work...
# - Write code
# - Create files
# - Install dependencies
# - Use subagents if stuck
# Mark as complete
~/.claude/skills/tasks/tasks.sh done 1
Use subagents when:
debugger - Stuck on errorsweb-researcher - Need documentation/examplescode-reviewer - Want quality checkworker - Delegate independent subtaskFor code changes:
# Run tests
npm test
pytest
go test ./...
# Check output
# Fix any failures before proceeding
For UI changes (MANDATORY):
~/.claude/skills/vision/vision.sh verify http://localhost:8000 "expected elements"
Call the QA agent to verify test coverage:
Use the qa-agent subagent to verify:
- All requirements tested
- UI verified (if applicable)
- Edge cases covered
QA will return:
QA VERIFIED → Proceed to supervisorQA GAPS FOUND → Add tasks for missing tests, loop back# Call supervisor to evaluate completion
# Supervisor will check everything and decide
Supervisor responses:
ALL COMPLETE → STOP, task is doneCONTINUE → More tasks remain, keep workingRE-PLAN → Requirements changed, start overCRITICAL: Never announce "done" yourself. Only stop when supervisor says ALL COMPLETE.
| Phase | Command | Required? |
|---|---|---|
| Start | /workflow start "desc" | Yes, for all tasks |
| Plan | Think, research, design | Yes |
| Tasks | /tasks add "..." | Yes |
| Work | /tasks working N + implement | Yes |
| Test | Run tests, /vision verify | Yes for code/UI |
| QA | qa-agent subagent | Yes |
| Supervisor | /supervisor | Yes, always last |
# 1. START WORKFLOW
~/.claude/skills/workflow/workflow.sh start "Build a web-based todo app"
# 2. PLAN (thinking)
# - Need: HTML, CSS, JavaScript
# - Features: Add, complete, delete todos
# - Storage: localStorage
# - Testing: Manual + vision verify
# 3. CREATE TASKS
~/.claude/skills/tasks/tasks.sh add "Create HTML structure"
~/.claude/skills/tasks/tasks.sh add "Add CSS styling"
~/.claude/skills/tasks/tasks.sh add "Implement JavaScript logic"
~/.claude/skills/tasks/tasks.sh add "Add localStorage persistence"
~/.claude/skills/tasks/tasks.sh add "Test functionality"
~/.claude/skills/tasks/tasks.sh add "Visual verification"
# 4. WORK ON EACH TASK
~/.claude/skills/tasks/tasks.sh working 1
# ... create HTML ...
~/.claude/skills/tasks/tasks.sh done 1
~/.claude/skills/tasks/tasks.sh working 2
# ... add CSS ...
~/.claude/skills/tasks/tasks.sh done 2
# ... continue for all tasks ...
# 5. TEST
# Start server, test manually
python3 -m http.server 8000 &
# Visual verification (REQUIRED for UI)
~/.claude/skills/vision/vision.sh verify http://localhost:8000 "todo input, add button, task list"
# 6. QA
# Call qa-agent to verify coverage
# 7. SUPERVISOR
# /supervisor will check everything and declare complete
| Mistake | Why It's Wrong | Correct Approach |
|---|---|---|
Skipping /workflow start | No tracking | Always start workflow for tasks |
| Writing code before planning | Miss requirements | Think first, code second |
Not using /tasks | Can't track progress | Break everything into tasks |
Skipping /vision verify for UI | UI bugs missed | ALWAYS verify UI visually |
| Announcing "Done!" to user | Only supervisor decides | Call /supervisor instead |
| Skipping QA | Test gaps missed | Always run qa-agent |