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?

Operation Type Validation Mechanism

Code changes Run tests, type check, lint

File operations Verify file exists, check content

API interactions Check response status, validate data

Build operations Build succeeds, no errors

Step 3: Design the Resolve Path

What happens on failure?

• Analyze: Parse error output to understand failure

• Fix: Make minimal, targeted corrections

• Re-validate: Run validation again

• 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

• Start with tests: If tests do not exist, consider adding them first

• Fast validation first: Run quick checks before slow ones

• Structured output: Use JSON for machine-parseable results

• Clear failure messages: Include context for resolution

• 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