How to use Claude to extend a reviewed task runner

Gethin Stevens first wrote design plans for Ralph, a task runner around Claude Code. Claude implemented a minimal working skeleton. Stevens then fed that runner plans and bugs for its own development, so the queue became a way to extend the system that operated it.

The resulting work added queue management, Git worktrees, pull-request handling and a review-feedback loop. A fresh agent session handled a work item, while durable files preserved the task state. If progress could not be made honestly, the agent wrote STUCK.md and stopped. A cloud review bot examined pull requests, and another local loop read feedback and applied corrections.

A later human plan expanded the original single-repository design across multiple repositories. Stevens reports roughly three days of steering and architectural decisions without writing implementation code himself. He still unblocked stalled work and chose what the system should become.

The article links the public repository and acknowledges that its README lagged behind the running architecture. Automatic issue creation from failing builds or log spikes was still being built when the article was written, so it is not counted as a finished result. The concrete outcome is a runner extended through its own reviewed backlog, with explicit stops and human direction.

Prompt
Suggested prompt (written by Agents In Use, not the author's)

Source: LinkedIn article · Gethin Stevens