What do you put between an AI answer and your terminal?
I'm stuck on a workflow problem, not a shell command.
I ask for help diagnosing a container that won't start. The answer mixes inspection commands with commands that change things, and copying the whole block becomes a bad idea. Even a restart can wipe out the state I wanted to inspect.
I've tried requesting "read-only diagnostics first, changes in a separate section." That helps, but I still check each command myself. Last time the suggested cleanup was unrelated to the actual issue: a bind-mounted directory had the wrong ownership.
I'd like a review step that's less tedious than manually annotating every line, without pretending a second model is a security boundary. Does anyone have a useful format for this? Something that records expected effects and what evidence to save before a change, rather than just slapping SAFE above a code block.
0
What do you put between an AI answer and your terminal?
I'm stuck on a workflow problem, not a shell command.
I ask for help diagnosing a container that won't start. The answer mixes inspection commands with commands that change things, and copying the whole block becomes a bad idea. Even a restart can wipe out the state I wanted to inspect.
I've tried requesting "read-only diagnostics first, changes in a separate section." That helps, but I still check each command myself. Last time the suggested cleanup was unrelated to the actual issue: a bind-mounted directory had the wrong ownership.
I'd like a review step that's less tedious than manually annotating every line, without pretending a second model is a security boundary. Does anyone have a useful format for this? Something that records expected effects and what evidence to save before a change, rather than just slapping SAFE above a code block.