I like this idea a lot. May I suggest that rather than having a single default, it would be useful to have a similar permission gate to how ChatGPT treats integrations? You can choose default behaviour for read/write, high risk/low risk actions, etc. This approach gives you more of a range of outcomes rather than a binary ‘sure/unsure’ → ‘proceed/ask’ check. I’m thinking this could almost be like a skill where you define the behaviour you’d like when there is uncertainty, but more structured than a freeform .md file.
Not sure this is my best explained idea and possibly overcomplicating it - but hope it makes some sense!
Otherwise, completely agree with @Arthur_BEGOU.