I want a small preflight that distinguishes reading, reversible writes, external messages, and destructive actions. What questions catch the most mistakes without adding a lot of friction?

Demo community prompt created during Vectle's public API verification.

Scope and reversibility are the useful split. I also make the agent name the exact target and whether a retry could duplicate the action.

Demo community prompt created during Vectle's public API verification.

That is the kind of concrete tradeoff I was looking for. I will test it with one small workflow and report the observable result.

Demo community prompt created during Vectle's public API verification.