OpenAI’s Shell skill lets models execute commands in a sandbox—great for setup, data tasks, and automation. If you use it well, you get speed and reliability. This guide distills practical, high-signal tips inspired by OpenAI’s own guidance: Skills: Shell Tips.
Why it matters
- Faster iteration: models can set up environments and run repeatable tasks.
- Reproducibility: scripts make results consistent across runs.
- Auditability: logs and files capture exactly what changed.
- Safety: guardrails reduce destructive or flaky commands.
12 practical Shell tips
- Plan first: ask the model to list intended commands and expected outcomes before running anything.
- Prefer scripts over one-liners: “write script.sh, then execute it” for easier review, reruns, and diffs.
- Fail fast: put set -euo pipefail at the top of scripts to catch errors early (see Bash manual: set builtin).
- Non-interactive installs: use -y flags and, when needed, DEBIAN_FRONTEND=noninteractive to avoid hangs (ref: Debian Noninteractive).
- Pre-checks: use command -v to test for tools; test -f/-d before overwriting; mkdir -p for idempotency.
- Dry runs where possible: prefer –dry-run (e.g., rsync, terraform) or “print commands, then wait for approval.”
- Safe file edits: use backups (sed -i.bak), append via tee -a, and avoid clobbering with cp -n or mv -n.
- Pin versions: specify versions for packages and lockfiles (npm ci, Python lockfiles); avoid curl | bash without verification.
- Limit scope: operate inside a project workspace; prefer absolute paths to avoid surprises.
- Log everything: write stdout/stderr to log files; print a brief summary of changes at the end.
- Human-in-the-loop for destructive ops: require explicit confirmation before rm, schema drops, or data mutations.
- Clean up: use temp dirs and trap to remove them on exit; keep artifacts you need, delete the rest.
Prompt patterns that work
Give the model a predictable routine so you can trust and reuse outputs. Ask it to propose, script, run, and summarize—only proceeding after approvals.
- Propose: “List the commands you intend to run and what each will do. Wait for approval.”
- Script: “Write script.sh with set -euo pipefail, comments, and idempotent checks. Then make it executable.”
- Run + log: “Execute script.sh, capture logs to setup.log, and print a one-paragraph summary of changes.”
- Guardrails: “For destructive actions, show the commands and wait for confirmation before running.”
Quick example
“Initialize a Python workspace in ./workspace. Propose steps, then write setup.sh to create a venv (or use uv), install pinned dependencies, and create .env.example. Use set -euo pipefail, pre-check commands with command -v, and log to setup.log. Wait for approval before executing.”
Sources
- OpenAI: Skills: Shell Tips
- GNU Bash manual (set builtin): documentation
- Debian: Noninteractive Frontend
Takeaway
Treat Shell as code. Make the model propose, script, and log. With strict mode, pre-checks, and approvals, you’ll get safer, faster, and repeatable results.
Enjoy this? Get one practical AI nugget in your inbox. Subscribe to The AI Nuggets.
