[VS Code] Implement Dialogue Engineering principles for AI coding assistants

Dialogue Engineering - The Best Way to Work With Your AI In November last year I took Jeremy Howard's Solve It course. The course demonstrated Jeremy's new way of working with AI coding assistants, that he dubbed "Dialogue Engineering". Dialogue engineering leans heavily into the idea that building software in the future will be a conversation between the engineer and the LLM. He also borrowed a lot of good ideas from the Mathematician Polya's book "How To Solve It", which is gives a timeless approach to problem solving. Unlike most of Jeremy's other courses which teach the fundamental building blocks of practical machine learning, this course aimed to teach students how to code by conversing with a LLM. It leveraged python's interpreted nature to allow both the user and the LLM to switch between chatting with each other and writing code. But this wasn't just a course on how to use existing LLM tools. Instead Jeremy leveraged his experience in building and researching in LLM's to transform an under-rated insight into a new interface paradigm. That LLM's tend to try to be consistent with whatever has been provided in their prompts. This is often at odds with the user, as if the LLM starts saying things that are wrong, or just misguided, then the user can only either try to keep talking or delete the response and try again. Jeremy's new alternative was to edit what the LLM responded with to be what you would like to see. Initially this feels weird, like having a conversation with someone and then changing what they said to be what you wish they had said. But LLM's don't actually have feelings, and they will happily continue on in the conversation. And it has the surprising result of keeping the LLM much more in line with what you are after. As I get deeper into the AI coding tools, I find myself wishing I could recreate these Dialogue Engineering principles in things like VS Code. Dialogue Engineering really helps the user keep the full context of the model front and centre, and paints a hopeful future where engineers and LLM's work together to build software. I haven't tried Dialogue Engineering with the more recent LLM's (such as Claude Sonnet 4), but I'm optimistic that this approach might end up being the dominant way of working with AI. Especially when the LLM model is much better at coding than they were in November 2024.