Thanks I will give it another try because our use case is HTML/Markdown documents, not code, and this could be interesting. I'm just hesitant to trust an LLM to do replacements and your broken example really didn't help with my confidence. Even a 1% error rate wouldn't be worth it because if find/replace doesn't work you know it doesn't work and can feed that error back into the agent to fix it (like how Claude Code recovers from its editing errors)

edit: The example is still broken. I've inspected the network request and it's definitely your backend that is broken not something being hardcoded... The CSS is not present in the request at all, but in the response it's being inserted.

took a closer look, the update snippet provided is referencing css thats not defined. this is sort of a overreach example of the "semantic" nature of the edits - ie, if an update explicitly contains syntax, or minor reference errors, the apply model corrects them. you can argue that this is an overstep in this example - but at the same time its not something claude/gemini would suggest

> the update snippet provided is referencing css thats not defined

That isn't how HTML works.

You mean the new HTML has classes which from what your LLM can see has no styles applied.

Maybe those styles are in a different file or maybe the developer didn't want any styles applied, the editing LLM would not know that.

It is completely broken behavior and you lied about it at first. Your model should not be writing code! It should just be applying edits given to it.