How to Write Shell Scripts with AI

A useful prompt describes the operating system, shell, input, expected output, and what must never happen. For example, specify whether the script is Bash on Linux or PowerShell on Windows, whether paths can contain spaces, and whether the task should only report changes or apply them. Ask for assumptions and a line-by-line explanation. This gives you a reviewable draft rather than an opaque command. Start with read-only behavior. A listing, count, or preview is easier to inspect than a script that moves or deletes data. Add a dry-run option that prints the exact operation for each item. Test with a temporary directory containing ordinary names, spaces, empty folders, hidden files, and a deliberately missing input. Check the output against what you expect before changing the script to perform the operation. Pay special attention to variable expansion, quoting, wildcard patterns, current working directory, permissions, and error handling. A path assembled from user input can point somewhere unintended. A command that works for one filename may break when a name contains spaces or punctuation. Require explicit input paths instead of relying on whichever directory happens to be current. For destructive operations, add a clear confirmation and refuse to run against a root, home, or otherwise broad path unless that behavior is intentional. Shell interpretation changes risk. Python's official subprocess documentation explains that argument lists avoid implicit shell interpretation, while explicitly invoking a shell makes the caller responsible for safely quoting metacharacters. Similar care applies when a generated script passes text into another command interpreter. Never paste an unexplained one-liner into an administrator terminal. Run with the least privilege needed, keep a backup, and review the diff or file list after execution. AI can also invent flags, assume a utility exists, or mix syntax from different shell versions.