Ask for a plan in plain language
Before a consequential command, ask the agent to identify the purpose, exact targets and expected changes. Ask it to separate inspection from mutation and local work from publication.
Keep secrets out of transcripts
Do not paste API keys, passwords, access tokens or private file contents into an explainer. Replace secrets with placeholders. Terminal output and command histories can preserve sensitive values.
Verify the outcome
A zero exit code alone does not prove the intended result. Check changed files, tests, destination records or remote state. Distinguish attempted, completed and verified actions.
Make approval specific
An approval to inspect is not approval to delete, publish or install. Name the exact action and scope. If the destination, cost, permissions or affected data changes, reconsider the decision.
Your next step
Pick one command from your agent’s last explanation. Identify its purpose, target and possible side effects before deciding whether to run it.