The problem with getting large language models to do what I want is that what I want is encoded in the words I use, and more importantly in the words I don't use. I don't use them because I failed to understand the problem clearly enough, I don't have enough domain expertise, or a variety of other reasons: assumptions, blind spots, laziness.

This isn't a prompt engineering problem — you can't prompt your way out of blind spots you don't know you have. It's not a context engineering problem either, and it's not just an Explore agent gathering more information for the model to read.

I've been experimenting with a dedicated agent step that takes my vague request, reads whatever context exists, and does the work of figuring out what I left unsaid. It produces something new — a synthesized version of what I actually meant, not just more context to read. Instead of the model interviewing me, it does that investigative work itself. I first called this an expander agent — it takes a compressed intent and expands it as though you had spent the time writing a detailed specification for whatever you want done.

I didn't invent this. But discovering it after trying to micromanage models, getting interviewed by them because I can't let go, and finally letting fatigue take over — that path led me here. I think it's useful enough that I wanted everyone else to know about it. The academic term seems to be intent decomposition, though what I'm doing is closer to intent expansion than the classical definition.

A bit more details about it here: https://labs.voidwire.info/posts/the-words-i-dont-use/

#llm #genai #claude