closed-loop-design（闭环设计）

Closed Loop Design Skill

Guide for designing closed-loop prompts that ensure agent reliability through feedback loops.

When to Use

• Designing new agentic workflows
• Adding validation to existing processes
• Creating self-correcting agent pipelines
• Building automated test-fix cycles

Core Concept

Every agentic operation should follow the Request-Validate-Resolve pattern:

REQUEST → VALIDATE → RESOLVE
    ↑                    ↓
    └────────────────────┘

Design Workflow

Step 1: Identify the Operation

What task needs to be validated?

• Code changes (feature, bug fix, refactor)
• File operations (create, modify, delete)
• External interactions (API calls, database queries)
• Build/deploy operations

Step 2: Define Validation Mechanism

How do we know if it succeeded?

Step 3: Design the Resolve Path

What happens on failure?

1. Analyze: Parse error output to understand failure
2. Fix: Make minimal, targeted corrections
3. Re-validate: Run validation again
4. Limit retries: Set maximum attempts (typically 3-5)

Step 4: Create the Prompt Structure

Template:

## Closed Loop: [Operation Name]

### Request
[Clear task description with success criteria]

### Validate
- Command: `[validation command]`
- Success: [what success looks like]
- Failure: [what failure looks like]

### Resolve (if validation fails)
1. Analyze the failure output
2. Identify root cause
3. Apply minimal fix
4. Return to Validate step
5. Maximum 3 retry attempts

Example: Test-Fix Loop

## Closed Loop: Implement Feature

### Request
Implement the user login feature according to spec.

### Validate
- Command: `pytest tests/test_auth.py -v`
- Success: All tests pass (exit code 0)
- Failure: One or more tests fail

### Resolve
1. Read failing test output
2. Identify which test failed and why
3. Fix the implementation (not the test)
4. Re-run pytest
5. Maximum 3 retry attempts

Common Closed Loop Patterns

Code Quality Loop

REQUEST: Write/modify code
VALIDATE: Lint + Type check + Tests
RESOLVE: Fix errors iteratively

Build Verification Loop

REQUEST: Make changes
VALIDATE: Full build succeeds
RESOLVE: Fix build errors

E2E Validation Loop

REQUEST: Implement user flow
VALIDATE: E2E test passes
RESOLVE: Fix failing steps

Best Practices

1. Start with tests: If tests do not exist, consider adding them first
2. Fast validation first: Run quick checks before slow ones
3. Structured output: Use JSON for machine-parseable results
4. Clear failure messages: Include context for resolution
5. Bounded retries: Prevent infinite loops

Memory References

• @closed-loop-anatomy.md - Full pattern documentation
• @test-leverage-point.md - Why tests are ideal validators
• @validation-commands.md - Validation command patterns

Version History

• v1.0.0 (2025-12-26): Initial release

Last Updated

Date: 2025-12-26 Model: claude-opus-4-5-20251101