Ralph Loop Plugin: Run Claude Code in a Self-Repeating Loop
ralph-loop is Anthropic's plugin for the Ralph Wiggum technique: a Stop hook re-feeds one prompt so Claude keeps iterating until a completion promise appears.
By Claudekit Team Published 1 min read
- Type
- Plugin
- Pricing
- Open Source
- Repository
- github.com
- Website
- github.com
ralph-loop is an Anthropic plugin that implements the Ralph Wiggum technique in Claude Code. You give it one prompt, and a Stop hook blocks Claude from exiting and feeds the same prompt back, so Claude keeps improving its own work in files until it outputs a completion phrase or hits an iteration limit.
Checked 24 Oct 2026 against the plugin’s README, hooks.json and marketplace.json.
Who it is for
Developers with well-defined tasks that have automatic verification, such as getting a test suite to pass, who are happy to let Claude iterate unattended.
Install
/plugin install ralph-loop@claude-plugins-officialStart a loop:
/ralph-loop "Build a REST API for todos. Requirements: CRUD operations, input validation, tests. Output <promise>COMPLETE</promise> when done." --completion-promise "COMPLETE" --max-iterations 50Stop it early with /cancel-ralph.
What it does
- Registers a
Stophook that intercepts Claude’s exit and re-sends your original prompt. - The prompt never changes; progress lives in your files and git history, which Claude reads on each pass.
--completion-promiseends the loop on an exact string match.--max-iterationscaps the number of passes.
Use cases
- Drive a feature to green tests with a TDD-style prompt.
- Greenfield prototypes you can leave running.
- Phased tasks where each phase has a clear, checkable outcome.
Pros
- No external scripts; it runs inside your session.
- Simple mental model: one prompt, repeated.
- Works well when tests or linters can judge success.
Cons and caveats
- Unlimited by default. The README recommends always setting
--max-iterationsas your main safety net. - The completion promise is a single exact string, so you cannot signal different outcomes like success versus blocked.
- Long loops consume a lot of usage.
Links
- Plugin folder: plugins/ralph-loop
- Related: hooks guide, Hookify, Claude Plugins Official
Frequently asked questions
How does the loop stop?
When Claude outputs the exact --completion-promise text, when it reaches --max-iterations, or when you run /cancel-ralph. Without --max-iterations the default is unlimited.
Does it run outside my session?
No. The loop happens inside your current Claude Code session through a Stop hook, so you do not need an external bash loop.
Does it work on Windows?
The stop hook is a bash script that needs Git for Windows. The README documents a workaround when bash resolves to WSL instead of Git Bash.
When is it a bad fit?
Tasks needing human judgment or design decisions, one-shot operations, unclear success criteria and production debugging.