ralph-loop

Automate iterative self-improvement by looping your task prompt back into the agent after each turn, feeding previous outputs into subsequent iterations to refine results. The ralph-loop skill executes this deterministic workflow by writing a scratchpad state file to track progress for resume and audit.

Set up state persistence for iterative development loops by creating a .cursor/ralph/ directory and writing a scratchpad.md state file. This scratchpad tracks iteration state automatically, enabling you to resume automated refinement loops and audit previous outputs.

You can control prompt feedback loop duration by configuring max_iterations and a completion_promise. These guardrails enforce deterministic termination, stopping the iterative refinement process once limits are reached or the promise condition is met.

The best way to stop iterative refinement early is by defining a completion_promise when starting the loop. The ralph-loop exposes a stop hook that enforces guardrails, automatically terminating the sequence once the promise condition evaluates successfully.

Iterative self-referential development works without external dependencies because the skill operates entirely within the agent environment. It creates local .cursor/ralph/ directories and manages state internally via a scratchpad.md file, requiring no external packages.

You need a scratchpad file for repeated agent task refinement to maintain state persistence across iterations. The scratchpad.md file logs each turn's output, ensuring the agent receives accurate prior context when feeding the prompt back into subsequent self-improvement cycles.