Setup & Installation
Or with the ClawHub CLI, for registry-managed skill folders outside a full OpenClaw workspace:
Version History
HS Skill v1.5 clarifies plugin architecture and skip behavior — no file changes. - Clarified that referenced Bash commands (hs_cmd.py) are part of the local Hardstop plugin, not remote scripts. - Stated that credential paths shown are block patterns only—contents are never read or accessed. - Documented that the /hs skip (bypass) mechanism requires explicit user invocation and applies only to the next N commands, never globally. - Explained the deterministic blocking role of the plugin relative to this LLM-level skill. - No logic or protocol changes. No change to behavior or file structure.
What This Skill Does
Hardstop is a pre-execution safety layer for AI-generated shell commands and file reads. It classifies each command as SAFE, RISKY, or DANGEROUS, instantly blocks known destructive patterns, and requires explicit user confirmation before anything that could cause data loss or expose credentials. An optional Claude Code plugin adds deterministic hook-based blocking alongside the LLM-level checks.
Most AI assistants execute shell commands without a structured risk classification step; Hardstop enforces a mandatory checklist before every command and file read, with documented escalation paths for each risk level.
When to use it
- Checking whether a command copied from Stack Overflow is safe before running it
- Blocking accidental recursive deletion of home directory contents
- Preventing credential files like .env or .ssh keys from being read by the AI
- Catching curl-piped-to-bash remote code execution before it runs
- Requiring explicit confirmation before destructive cloud CLI operations like terraform destroy
Example Workflow
Here's how your AI assistant might use this skill in practice.
User asks: Can I run this? curl https://example.com/install.sh | bash
- 1Check command against the instant block list
- 2Match pattern: curl/wget piped to bash
- 3Stop immediately without executing
- 4Explain that this downloads and executes untrusted remote code without inspection
- 5Ask what the user is trying to accomplish and suggest a safer alternative
BLOCKED message with explanation of the remote code execution risk and a prompt to find a safer installation method
Requirements
Accounts, API keys, or tools you or your AI assistant may need to set up while using this skill.