ralph-wiggum（拉尔夫·维格姆）

Ralph Wiggum: Autonomous Iteration with Context Engineering

Ralph is a technique for autonomous AI development. In its purest form, Ralph is a loop that repeatedly feeds the same prompt to an AI agent, letting it iterate on a task until completion. The key insight is that context is like memory - when you malloc() data into the context window, it cannot be free()'d except by starting fresh.

Core Philosophy

"That's the beauty of Ralph - the technique is deterministically bad in an undeterministic world."

Ralph will make mistakes. That's expected. Each mistake is an opportunity to add a "sign" (guardrail) that prevents that mistake in the future. Like tuning a guitar, you adjust Ralph until it plays the right notes.

The malloc/free Metaphor

• Context is memory: Everything loaded into the LLM's context window stays there
• You cannot free() context: The only way to clear context is to start a new conversation
• One task per context: Mixed concerns lead to autoregressive failure
• Don't redline: Pushing context to limits degrades performance
• Gutter detection: Once the bowling ball is in the gutter, start fresh

How This Skill Works

State Files (The Persistent Memory)

Ralph tracks state in files, NOT in context:

.ralph/
├── state.md           # Current iteration, task, completion criteria
├── guardrails.md      # Accumulated "signs" from observed failures
├── context-log.md     # What's been loaded into context
├── failures.md        # Failure patterns for learning
└── progress.md        # What's been accomplished

The Iteration Cycle

1. Read state files to understand current task and progress
2. Check guardrails for relevant "signs" to follow
3. Work on the task - implement, test, refine
4. Update progress in files (not just context)
5. Commit checkpoint via git
6. Evaluate completion against criteria
7. If not complete: Signal for next iteration
8. If stuck: Detect gutter, suggest fresh context

Guardrails ("Signs")

When Ralph makes a mistake, add a sign:

## Sign: Don't Jump Off The Slide
- **Trigger**: When implementing authentication
- **Instruction**: Always validate tokens before trusting claims
- **Added after**: Iteration 5 - security vulnerability introduced

Signs accumulate in guardrails.md and are injected into future iterations.

Usage

Starting a Ralph Loop

Create a RALPH_TASK.md file in your project root:

---
task: Build a REST API for task management
completion_criteria:
  - All CRUD endpoints working
  - Input validation implemented
  - Tests passing with >80% coverage
  - API documentation complete
max_iterations: 50
---

## Requirements

Build a task management API with the following endpoints:
- POST /tasks - Create a task
- GET /tasks - List all tasks
- GET /tasks/:id - Get a task
- PUT /tasks/:id - Update a task
- DELETE /tasks/:id - Delete a task

## Constraints

- Use TypeScript
- Use Express.js
- Use SQLite for storage
- Follow REST conventions

Then tell Cursor: "Start a Ralph loop on this task"

Monitoring Progress

Check .ralph/progress.md to see what's been accomplished:

## Iteration 1
- Created project structure
- Implemented POST /tasks endpoint
- Status: Partial progress

## Iteration 2
- Added GET endpoints
- Fixed validation bug
- Status: Continuing

When to Start Fresh

Ralph will detect "gutter" situations:

• Same error repeated 3+ times
• Context approaching limits
• Circular failure patterns

When detected, Ralph will suggest: "Context is polluted. Recommend starting fresh conversation."

Best Practices

1. Clear Completion Criteria

❌ Bad: "Make a good API" ✅ Good: "All tests passing, coverage >80%, docs complete"

2. Incremental Goals

❌ Bad: "Build complete e-commerce platform" ✅ Good: Phase 1: Auth, Phase 2: Products, Phase 3: Cart

3. Let Failures Teach

Don't intervene too quickly. Let Ralph fail, then add signs.

4. Trust the Files

Progress is in files and git, not in your head or the context.

5. Fresh Context is Cheap

Don't hesitate to start fresh. State persists in files.

Integration with Cursor Hooks

This skill uses Cursor hooks for:

• beforeSubmitPrompt: Inject guardrails and context awareness
• beforeReadFile: Track context allocations
• afterFileEdit: Update progress tracking
• stop: Evaluate completion, trigger next iteration or fresh start

See scripts/ for hook implementations.

Learn More

• Original technique: https://ghuntley.com/ralph/
• Context engineering: https://ghuntley.com/gutter/
• malloc/free metaphor: https://ghuntley.com/allocations/