CLAUDE.md vs START_HERE.MD

I like to use other harnesses other than Claude from time to time; because I want to burn tokens but not money, and I like to write prompting experiments that are measurable, which means they have potential to get better!

So for something like this, I like to use Opencode CLI instead of Claude CLI.

And I noticed that sometimes some models don't respond well to CLAUDE.md as an entry point. So, I went through several iterations using auto researcher. And it looks like START_HERE is doing pretty well. Better than CLAUDE and even README which surprised me. I'm not saying this a gold standard but wanted to share.

6 comments