FWIW all my coding with LLMs is very hands-on. What I've ended up doing with LLMs is something like the following: 1. New conversation. Describe at a high lev

FWIW all my coding with LLMs is very hands-on. What I've ended up doing with LLMs is something like the following: 1. New conversation. Describe at a high level what change I want made. Point out the relevant files for the LLM to have context. Discuss the overall design with the LLM. At the end of that conversation, ask it to write out a summary (including relevant files to read for context next time) in an "epic" document in llm/epics/. This will almost always have several steps, listed in the document. Then I review this and make sure it's in line with what I want. 2. New conversation. We're working on @llm/epics/that_epic.md. Please read the relevant files for context. We're going to start work on step N. Let me know if you have any questions; when you're ready, sketch out a detailed plan of implementation. I may need to answer some questions or help it find more context; then it writes a plan. I review this plan and make sure it's in line with what I want. 3. New conversation. We're working on @llm/epics/that_epic.md. We're going to start implementing step N. Let me know if you have any questions; when you're ready, go ahead and start coding. Monitor it to make sure it doesn't get stuck. Any time it starts to do something stupid or against the pattern of what I'd like -- from style, to hallucinating (or forgetting) a feature of some sub-package -- add something to the context files. Repeat until the epic is done. If this sounds like a lot of work, it is. As xkcd's "Uncomfortable Truths Well" said, "You will never find a programming language that frees you from the burden of clarifying your ideas." LLMs don't fundamentally change that dynamic. But they do often come up with clever solutions to problems; their "stupid questions" often helps me realize how unclear my thinking is; they type a lot faster, and they look up documentation a lot faster too. Sure, they make a bunch of frustrating mistakes when they're new to the project; but if every time they make a patterned mistake, you add that to your context somehow, eventually these will become fewer and fewer.