[softwaredoug]

There is some joy in catching the LLM in a house of cards misunderstanding of how something works. They’ll weave really convincing fictions about from a bit of debugging. Then the Socratic probing of that can be fun. In part because the LLM tries to use big words to often seem smart.

Then you hopefully capture that information somehow in a future prompt, documentation, test, or other agent guardrail.

So I finds fun in the knowledge engineering of it all. The meta practice of building up a knowledge base in how to solve problems like this codebase.