Show HN: A scriptable text editor for LLMs
github.comBuilt this out of curiosity to see whether LLMs get better a editing if they express edits as a series of operations instead of a diff.
I don't have my answers yet, but actually building a PoC was the first step.
Claude seems to be happy to use the tool and write fake elisp just fine.
Think of it as sed without the `s`, and a tiny bit of Emacs instead.
What I'm looking for: fellow experimenters to get feedback and maybe develop the idea further.
The implementation is intentionally naïve.
Have fun!
It's certainly worth exploring. Thee way most agents work right now bugs me. Sending huge diffs seems inefficient, uses a lot of tokens, and is prone to error. Another option that came to mind is working at the AST level, but that's a whole other can of worms to open.