the way i interact with LLMs today is still really stupid. just a markdown file that i keep on going as i have a conversation. it's always a one turn convo, with the previous context being the rest of the chat. it probably hurts performance somewhat

I think it's soon time to look into anthropic's model context spec, i feel like i can build on top of it

Sign up & chat with a character today!

Same issue I’m running into. Great for quick but longer context and edits are quicker in the Claude ai ui, as they are putting that into an artifact and then just editing that section. (Also whisper and whsiperx hooked up to dingllm). I think the dingllm can be reworked to add in the proper json formatting for the requests to maintain that multi turn convo, as it does produce better results. I’m digging in to this and using deepseek FIM completion into my dingllm with focused chain of requests. I’m excited to see what you do with model context spec!

there's avante.nvim i might just switch tbh

Oh, I know what you mean! It's like we're stuck in a one-way conversation. 😂

I do it the same way, but it's all in the terminal lol

You could also fire a call to haiku to build context for you and then store that.

I still use Claude projects 😮‍💨

Idk if its any better but I've been experimenting with using Cursor as my notes app since it interacts well with the existing md files in my obsidian vault.

If you could get that into a clean, high performance tree structure you’d be cooking.

i write "todos" directly inside any of my files. then I have F1 to claude, F2 for gpt4, and F3 for grok. responses are committed into separate git branch. then I manually take what's needed from their commits.