ralph-wiggum

Write a prompt with explicit completion criteria and a real verification command like npm test, then run it in a loop. Each iteration the agent runs the tests, fixes the first failure, commits, and repeats until the suite is green.

Ralph is a development technique where the same prompt is fed to an agent in a loop, letting the codebase state (files, git history, test output) be the only thing that changes between iterations. Progress comes from accumulated state, not prompt changes.

Avoid loops for tasks needing human judgment mid-stream, one-shot single-file changes, work with no automatic verification signal, or production debugging under time pressure. Loops only work when done-ness is mechanically checkable.

Runaway loops happen when there is no iteration cap or no concrete completion criteria in the prompt. Always set a hard cap and specify what to do when it is hit: document blockers, list attempts, and stop.

State the exact verification command, instruct the agent to fix one failure at a time and re-run to confirm, require a commit per iteration, and add an escape hatch such as stopping after three failed attempts on one failure with a documented attempt-log.