I used an LLM to review what changes the likelihood of a PR getting merged. I then asked it to select potential changes based on its findings. I selected from those options and carefully guided the LLM to produce several pull requests. I fed the subset of PRs that failed back into the skill and discovered additional missing factors.