One-shot performance of LLMs is a good measure of many of its capacities, but it is only sufficient for tasks that can be objectively and automatically verified. I believe that in most real-world cases this is not the case: people have vague, unquantifiable, taste-based, complex goals, and it's the way one works closely with AI over many prompts that makes the difference between slop and actual work. But few people seem to talk about that today. This blog post (and the browsable prompt trascripts it links to) is an attempt at changing that.

I'm sharing all my prompts for a little toy web-app project. It's the opposite of one-shot! I also (over?)analyzed and charted my data. I'd love to hear everyone's thoughts, not so much on the exact execution (debatable) but on my proposal to talk more about HITL AI workflows.