I'd never let a junior dev use one of these models, but I've been coding for 30 years and know how to catch mistakes. It saves a huge amount of time.

I've done this multiple times and Claude explicitly provided a config parameter in the script to make it "read only" by default.

Regardless, why wouldn't anyone test it on a small subset of your photos before trying it on a full collection? You would do it with a script you wrote personally and you should do it with an LLM script as well.