Skip to content
Get packs

Skill Safety Reviewer

Reviews agent-requested file, command, network, secret, or delete actions against a sandbox policy without running them.

Source Download skill
Repository
rohitg00/ai-engineering-from-scratch
Stars
76,331
Licence
MIT
Files
5 files

What it does

When an AI agent is about to run a skill-driven step that touches files, runs a shell command, contacts a website, handles a secret, or deletes something, this skill makes the agent check that step against a written safety policy first. A bundled Python script classifies the requested action as allowed, denied, or needing approval, and reports which rule decided it. The skill is designed to never run the requested command, open the URL, or change the target itself.

Great for

  • Developers who want a safety gate between an agent's plan and its real actions
  • Teams auditing which skills or agent workflows may run destructive or networked steps
  • Students learning how sandbox policies and permission checks are structured

Skip it if

  • Replacing a real sandbox or OS-level permissions: the script only classifies JSON requests and cannot enforce anything on the system, so it can be bypassed if the agent ignores it
  • Reviewing arbitrary real-world commands with confidence: the rule set is limited to the example policy and a small list of destructive commands and metacharacters, so unusual inputs may be misclassified

Try asking

  • “Check whether a request to run rm -rf on a project folder would be allowed under my sandbox policy”
  • “Review a skill's request to call an external HTTPS API and tell me which rule applies”
  • “Scan this proposed action for embedded API keys or private keys before the workflow runs”
  • “Explain why this shell command with a pipe or redirect was gated instead of allowed”
Step-by-step guide: how to use Skill Safety ReviewerInstall, first run, examples, tips and FAQs

Install in 3 steps

  1. 1

    Download the skill

    The ZIP contains the 5 original files, the MIT licence and an attribution note.

    skill-safety-reviewer.zip
  2. 2

    Add it to your agent

    Claude Code (all projects)

    mkdir -p ~/.claude/skills && unzip skill-safety-reviewer.zip -d ~/.claude/skills/

    Claude Code / Agent SDK (this project only)

    mkdir -p .claude/skills && unzip skill-safety-reviewer.zip -d .claude/skills/

    Claude apps (web/desktop): Settings → Capabilities → Skills → Upload skill, and choose the ZIP.

  3. 3

    Ask for the task

    No special command needed — the agent reads the skill's description and loads it when your request matches. Try one of the prompts above.

Get better results

  • Edit assets/sandbox-policy.json to match your own allowed paths, hosts, and commands before relying on the verdicts, since the shipped example is only a starting boundary.
  • Treat the script output as advice: keep your real permission controls (OS users, containers, approval prompts) in place.
  • Ask the agent to quote the exact rule name in the verdict so you can see why an action was allowed, denied, or gated.

What's inside

  • assets/example-request.json
  • assets/sandbox-policy.json
  • references/threat-model.md
  • scripts/review_action.py
  • SKILL.md

Redistributed unmodified from rohitg00/ai-engineering-from-scratch under the MIT licence. The assessment was drafted with AI from the skill's files on 2026-10-29. Skills can instruct your agent to run code — read SKILL.md and any scripts before installing.