Post

For better prompts and evals, instead of just asking for reasoning, I'm telling the LLM to give me prompt feedback directly. Feels more useful for error analysis, like it helping me help it. Then I run it on a bunch of samples to see where we don't align.