I’ve been reading a lot lately about making it clearer what prompts are used to elicit certain outputs when coding with agents.
Cloudflare’s OAuth library, written largely by LLM agents with human review, includes the prompts in the PR comments, making it much easier to figure out the author’s intent and imagine how you might prompt an agent to build something similar yourself.
I feel there needs to be a better way to automatically commit after each agent loop is executed, including the prompt as part of that commit. From there, we can fluidly move back and forth between commits as necessary, accepting or tossing the resulting output from the agent.
When working on a feature, you create a new branch, then prompt the agent as many times as needed, and the agent makes a commit after each loop is finished. Once you’re satisfied with the results, you can squash and merge, preserving all the prompts, or merge everything to main if you prefer.