We no longer write any of the prompts in our codebase

If you are anything like us, your codebase probably has some prompts in it. What we have observed is that LLMs can auto-improve them significantly. The key insight is that, to go beyond few-shot prompting, a thinking agent can now iteratively evaluate the quality of results, hypothesize improvements, and test them out until converging toward the optimal prompt.

We’ve been using this heavily in our work and have got great results in terms of user satisfaction and the relevance of our system.

I’m sharing here with you the protocol we use: you can just give the link to your coding agent and it will know how to execute it.

Hope this will be useful to at least some of you. And happy to chat about how you are optimizing prompts in your work?