Ralph: A Technique Review

Ralph is a technique named for a Simpson character, one Ralph Wiggiams, who will continue to refine something until an arbiter says it is successful.

Often the workflow goes as follows:

• Refine a specification (often referred to as a prd)
• Generate a context file like progress.json
• Do
  • Have an AI agent work against the prd until it hits it’s context limit. This is done in a fresh context every time.
  • Have an arbiter agent review progress.json to see if all items are done.
    • If not done, refine the progress.json file and repeat
    • If done then stop

Sources

• A Skills implementation of Ralph for Claude https://github.com/snarktank/ralph
  • https://raw.githubusercontent.com/snarktank/ralph/refs/heads/main/ralph-flowchart.png (https://github.com/snarktank/ralph/blob/main/ralph-flowchart.png)
  • Core Idea:
    • Fresh Context, only retaining: Git history, progress.txt, and prd.json
• Blog Entry: https://ghuntley.com/ralph/
• Block Goose support: https://block.github.io/goose/docs/tutorials/ralph-loop/