Thank you!

that's another option of course. But it's definitely easier to setup all these checks and linter tests with a skill file vs git hooks and actions

what danger?

Maybe like that the humans would not properly review the more complicated ones too? I don't think it's a big danger to be clear.

yeah I guess figuring out how AI and your team can optimally work together is not that straightforward.. probably every engineering team is trying to figure that out atm :D but if we already let AI write reviews, they should at least be as good as they can

fair point. I think they should ALSO be in your CLAUDE.md. Doing Claude reviews with the exact same instructions still makes sense because the PR itself is often a co-production between AI and the human who gives the instructions (and in the coding process the agent might be more sycophantic toward the user's ideas). Additionally, when writing code, Claude is more focused on writing code that works and puts a lot of your repo into context, while during reviewing it pays more attention to these guidelines

storing the classification definitely makes sense! We re-use the classification for the different merge attempts in the same run but do not store it because mostly we work with different data every time

So fuzzy matching only makes sense if you expect two columns having the same data more or less, otherwise you can skip that step.

And then you have to pick a threshold -> if similarity of strings is above that threshold, it's a match, otherwise, not. Threshold should be high to prevent false positives. LLM will take care of the non-matches