Ralph Loop

# Ralph Loop - Simple Autonomous Development

The simplest form of autonomous AI development. Named after Geoffrey Huntley's technique.

## The Core Loop

```bash
while :; do cat PROMPT.md | claude ; done
```

That's it. That's the tweet.

## How It Works

1. Write your task in `PROMPT.md`
2. Run the loop
3. Claude executes, fails, learns, retries
4. Eventually it succeeds
5. Kill the loop (Ctrl+C)

## Example PROMPT.md

```markdown
# Task: Add dark mode to the settings page

## Requirements
- Toggle switch in settings
- Persist preference to localStorage
- Apply theme class to document root
- Respect system preference as default

## Completion Criteria
When you're done:
1. Run `npm test` - all tests pass
2. Run `npm run build` - no errors
3. Say "TASK COMPLETE" at the end of your response

## Current Status
Check git status and continue from where you left off.
```

## Why It Works

- **Failures are information** - Each failed attempt teaches Claude what doesn't work
- **Tests as guardrails** - Automated validation catches mistakes
- **Fresh context each loop** - No context window bloat
- **Persistence wins** - Keep trying until it works

## Tips

1. **Clear completion criteria** - How does Claude know when to stop?
2. **Verifiable success** - Tests, builds, linters
3. **Incremental progress** - Commit after each successful step
4. **Watch the costs** - Each loop is an API call

## When to Use

- Tasks with clear pass/fail criteria
- Overnight unattended runs
- Iterative refinement (styling, tests, refactoring)
- Bug fixes with reproduction steps

## When NOT to Use

- Subjective decisions ("make it prettier")
- Production debugging
- Anything needing human judgment mid-task

## The Philosophy

> "Instead of trying to get it right the first time, let the machine iterate until it converges on a solution."

Stop prompt engineering. Start loop engineering.