A lot more tricks can be learned from just watching AI work. Instead of allowing AI to work autonomously, go back to the old days where you manually approve every command the AI runs. Just recently while doing performance optimization work, I found Opus using the `perf` command in ways I didn’t know possible. Just give AI a real task and carefully read what commands are used by the AI to solve the problem; most likel…
> go back to the old days where you manually approve every command the AI runs Those are also the current days if you have any sense. It's a bad idea to run an LLM with access to your machine at all, but if you absolutely must, you better review everything it does to make sure it doesn't run anything insane.
The rest of us picks a level between "approve some" and "YOLO" depending on how much they worry about having to clean up the workspace afterwards. Encouraging using the repo and some dedicated per-session storage, while discouraging changes to global state (like installing packages system-wide) reduces cleanup problem to "every now and then `rm -rf` some agent session dirs".