Damus
Mitch Downey :pci: · 5w
FWIW the process I have used for large tasks with LLM programming looks like: 1) Use Cursor. It behaves far less quirky than VS Code for programming in my experience. 2) Explain what I want in detai...
Mitch Downey :pci: profile picture
4) When it is confident, and you've reviewed the plans, and you're ready to go, ask it to implement each plan, one prompt per plan at a time, until it completes all the work.

There's more nuance I could get into, and some custom skill/rule files, but that's the general idea. Basically, I try to get the LLM to plan the work out thoroughly and in bite sized, deferred chunks before asking it to do the work. I haven't found the need to use expensive Anthropic models most of the time.
Mitch Downey :pci: · 5w
I don't love using these proprietary models btw. I'd love to rely on my own self-hosted and open source models, but afaik the cost of doing that without it being an absolute slog and fighting with inaccuracies is very high. For some reason GPT is much cheaper than Anthropic. I do think Anthropic mod...