Practical guides · 07
Preview a risky command before it runs.
Read the risk badge, see which files a delete or move would touch, and keep a backup step between you and an overwrite.
4 minute read
Most terminal mistakes that hurt are not typos. They are commands that did exactly what they said, to more files than you expected: a rm with a wildcard, a mv into the wrong folder, a chmod -R one level too high. AITerm gives every command, whether you typed it or the AI suggested it, a risk assessment before it runs. This guide shows how to use that assessment to look before you leap.
1. Read the badge.
Each command is shown with a risk badge:
- Safe (green) looks safe to run.
- Caution (orange) is worth a glance.
- Destructive (red) deletes or overwrites data.
- Irreversible (red) is the most severe tier: system-wide and cannot be undone.
Click the badge to open the risk popover. It lists the findings behind the tier, can offer a safer alternative with Use this instead, and has an Explain in detail option. A badge is an assessment, not a guarantee, so a green badge is still a reason to read the command, not to skip reading it.
2. Preview what it will touch.
For a risky command, the popover offers Preview what this will do. For a delete, move or chmod, you get a read-only list of the files it would touch, computed without running anything. For tools such as rsync that have their own dry-run form, AITerm can build it for you: choose Load dry-run to put that version in the input bar, then run it through the normal approval step and read its output.
If the list is longer than you expected, stop there. Narrow the path or the pattern, or ask in plain English for a more specific command, and preview again.
3. Back up, then run.
When a destructive command of your own reaches the confirmation alert and a backup makes sense, the alert can offer up to four choices:
- Run anyway runs the original now, with no backup.
- Back up, then run runs a backup first. The original runs only if the backup succeeds; if the backup fails, the original is not run and you are told so.
- Use safer alternative loads a suggested safer command into the input bar. Since 0.25.7 it arrives with its own risk badge, so read that badge before you approve it.
- Cancel does nothing.
The backup is its own command and passes the same safety check as anything else.
4. Practice with Dry Run mode.
If you are learning, or about to try a command you found online, turn on Start in Dry Run Mode in Settings ▸ Security. AITerm then evaluates and previews every command but runs nothing until you turn it off. A Dry Run pill in the status strip reminds you it is on, and the setting applies to every open tab at once. Turn it off when you are ready to run for real.
5. Tighten the rules when the stakes rise.
The AI ▸ Safety Profile menu switches between four profiles, and each one only ever makes the gate stricter. Personal is the default and asks for confirmation on Irreversible commands. Work also asks before destructive commands. Production asks before caution and destructive commands and blocks Irreversible ones. Locked Down asks before caution commands and blocks destructive and Irreversible ones. Switching to a stricter profile before you work on a server or a shared project is a quick way to add a pause where it matters. Afterward, AI ▸ Security Log… shows what ran, what was blocked and what was only previewed.
Try it in AITerm.
Risk badges, dry-run previews, backup prompts, Safety Profiles and the security log are part of the free core app, along with AI command suggestions, /fix, /explain, tabs, split panes, saved workspaces, local Ollama setup and supported bring-your-own-key connections. Runbooks and the autonomous agent are Pro. AITerm runs on Apple Silicon Macs with macOS 13 or later. Cloud provider charges are separate.
Keep learning.
- Three checks before you run an AI-generated command
- When a command fails: read it, explain it, then fix it
- Safety and security
Prepared with AI assistance from the current AITerm documentation. The steps follow the handbook; this is not a recorded app demonstration.