What's best way to handle AI agents that treat every prompt as an invitation for a massive refactor?
by•
I spent an hour reviewing a diff today where the agent decided my database schema needed a modern upgrade
even though I just wanted to add a simple validation check.
It’s exhausting having to babysit the diff just to make sure it didn't touch background logic.
3 views
Replies
What I do is put the scope into the standing instructions file instead of each prompt. Mine has lines like: never reformat or clean up code you were not asked to touch, do not add error handling or validation for cases that cannot happen, no comments unless the reason is not obvious. Several agent sessions share my repo, so there is also a rule to stage only your own files by name. With the rules in the file the agent reads every time, you stop repeating them per prompt, and reviewing a big diff gets a simple question: which rule did this break.