workflow-designer（工作流程设计师）

Workflow Designer

The Workflow Designer skill helps you design, document, and optimize multi-step AI-powered workflows. It applies process design principles to break down complex tasks into clear, executable sequences that leverage Claude Code's capabilities, external tools, and multi-agent coordination.

This skill guides you through workflow analysis, identifying optimal task decomposition, determining when to use automation versus human input, and documenting workflows in a format that's both human-readable and AI-executable. It helps you think through error handling, branching logic, validation steps, and integration points.

Use this skill when you're tackling complex, multi-step processes that could benefit from AI assistance, or when you need to transform ad-hoc procedures into repeatable, documented workflows.

Core Workflows

Workflow 1: Design New Workflow from Requirements

1. Clarify the goal:

- What outcome is needed? - Who initiates the workflow? - What triggers completion?

1. Identify inputs and outputs:

- What data/context is required to start? - What should the workflow produce? - What format should outputs take?

1. Map the process:

- Break into logical phases - Identify decision points - Map dependencies between steps - Note parallel vs sequential operations

1. Design error handling:

- What can go wrong at each step? - How should failures be handled? - What requires human intervention?

1. Assign responsibilities:

- Which steps are AI-executable? - Which require human judgment? - Which need external tools/APIs?

1. Document the workflow:

- Step-by-step instructions - Decision trees - Validation checkpoints - Success criteria

1. Test with sample scenario
2. Refine based on results

Workflow 2: Optimize Existing Workflow

1. Analyze current workflow:

- Map current steps - Identify bottlenecks - Note repetitive tasks - Find error-prone areas

1. Identify optimization opportunities:

- What can be automated? - What can be parallelized? - What steps are unnecessary? - Where are handoffs inefficient?

1. Redesign with improvements:

- Consolidate redundant steps - Automate repetitive tasks - Parallelize independent operations - Add validation early

1. Compare before/after:

- Time savings - Error reduction - Complexity changes - Resource requirements

1. Document changes and rationale
2. Plan migration from old to new
3. Test new workflow thoroughly

Workflow 3: Break Down Complex Task

1. Understand the complex task:

- What makes it complex? - What are the components? - What are the constraints?

1. Decompose into subtasks:

- Identify logical boundaries - Group related operations - Order by dependencies

1. Define interfaces:

- Inputs for each subtask - Outputs from each subtask - Data flow between tasks

1. Assign to appropriate executors:

- AI agents - External tools/MCPs - Human review points

1. Add coordination layer:

- How do subtasks communicate? - What monitors overall progress? - How are results integrated?

1. Document the decomposition
2. Validate completeness

Workflow 4: Add Error Handling & Resilience

1. Map potential failure points:

- External API failures - Invalid inputs - Resource constraints - Timeout scenarios

1. Design error handling strategy:

- Graceful degradation - Retry logic with backoff - Fallback options - Error reporting

1. Add validation checkpoints:

- Pre-conditions before steps - Post-conditions after steps - Intermediate result validation

1. Implement recovery mechanisms:

- State preservation - Resume from checkpoint - Rollback procedures

1. Document error scenarios:

- What to do when X fails - How to recover - When to escalate

1. Test failure scenarios

Quick Reference

Best Practices

• Start with Outcomes: Define success before designing steps

- What does "done" look like? - What artifacts should exist? - What state should system be in?

• Keep Steps Atomic: Each step should be single-purpose

- One clear action - One clear output - Easy to validate - Easy to replace or improve

• Make Dependencies Explicit: Show what relies on what

- Use directed graphs or numbered dependencies - Identify parallelizable operations - Note blocking dependencies

• Plan for Failure: Every workflow has edge cases

- What if API is down? - What if input is malformed? - What if process takes too long? - What if user cancels?

• Add Validation Early: Catch issues before they cascade

- Validate inputs before processing - Check outputs before passing to next step - Verify assumptions at decision points

• Document Decision Points: Make branching logic clear

- What triggers each path? - What are the criteria? - Who makes the decision (AI vs human)?

• Separate Concerns: Group related operations

- Data collection phase - Processing phase - Validation phase - Output generation phase

• Make It Resumable: Long workflows should support interruption

- Save state at checkpoints - Enable resume from last checkpoint - Track progress explicitly

• Test with Edge Cases: Don't just test happy path

- Malformed inputs - Missing dependencies - Timeout scenarios - Concurrent execution

Workflow Design Patterns

Sequential Pipeline

Input → Step 1 → Step 2 → Step 3 → Output

Use when: Each step depends on previous step's output Example: Data ingestion → Validation → Transformation → Storage

Parallel Execution

Input → [Step 1, Step 2, Step 3] → Merge → Output

Use when: Independent operations can run concurrently Example: Lint, Test, Type-check → Aggregate results → Report

Conditional Branching

Input → Decision → [Path A | Path B] → Merge → Output

Use when: Different paths based on conditions Example: File type detection → [JSON parser | CSV parser] → Normalize

Iterative Refinement

Input → Process → Validate → [Done? → Output | Refine → Process]

Use when: Output quality improves through iterations Example: Generate code → Review → [Acceptable? → Deploy | Fix issues → Generate]

Multi-Agent Orchestration

Input → Coordinator → [Agent A, Agent B, Agent C] → Synthesizer → Output

Use when: Complex task needs specialized sub-agents Example: Feature request → Planner → [Designer, Developer, Tester] → Integrator → PR

Event-Driven Workflow

Trigger → [Handler 1, Handler 2, Handler N] → Aggregate → Output

Use when: Workflow responds to events/webhooks Example: Git push → [Build, Test, Deploy, Notify] → Status update

Workflow Documentation Template

## Workflow: [Name]

### Purpose
[What this workflow accomplishes]

### Triggers
- [What initiates this workflow]

### Inputs
- [Required inputs]
- [Optional inputs]

### Steps
1. **[Step Name]**
   - Action: [What happens]
   - Owner: [AI | Human | Tool]
   - Input: [What this step receives]
   - Output: [What this step produces]
   - Validation: [How to verify success]
   - On failure: [What to do if this fails]

2. **[Step Name]**
   [...]

### Decision Points
- **[Decision Name]**
  - Condition: [What determines the path]
  - If true: [Path A]
  - If false: [Path B]

### Outputs
- [What the workflow produces]
- [Where outputs are stored/sent]

### Success Criteria
- [How to know the workflow succeeded]

### Error Handling
- [Common failures and responses]

### Estimated Duration
- [How long this typically takes]

### Dependencies
- [External tools/services required]
- [Other workflows this depends on]

Workflow Metrics to Track

When designing workflows, consider measuring:

• Duration: How long does the workflow take?
• Success rate: What percentage complete successfully?
• Bottlenecks: Which steps take longest?
• Failure points: Where do errors occur most?
• Retry rate: How often do steps need retrying?
• Human intervention: How often is manual action needed?
• Resource usage: What's the computational/financial cost?

Common Pitfalls

• Over-engineering: Don't add complexity for problems that haven't occurred
• Under-specifying: Don't assume steps are obvious without documentation
• Ignoring failures: Don't design only for happy path
• Tight coupling: Don't make steps too dependent on implementation details
• No rollback: Don't make destructive operations irreversible
• Missing validation: Don't pass bad data between steps
• Sequential when parallel works: Don't serialize independent operations
• No progress tracking: Don't make long workflows black boxes