guard
by garrytanguard is a full safety mode for Workflow Automation that combines destructive-command warnings with directory-scoped edit restrictions. Use the guard skill to guard against mistakes, lock down edits, and block risky commands like rm -rf or force-pushes when working near production or live data.
This skill scores 68/100, which means it is worth listing with caveats: it has a real, triggerable safety workflow, but users should understand that it depends on sibling skills and lacks broader adoption scaffolding. For directory users, it is a practical install candidate if they want enforced guardrails for destructive commands and directory-scoped edits.
- Clear trigger coverage for safety-related requests: "full safety mode", "guard against mistakes", and "maximum safety".
- Concrete operational behavior: hooks check Bash for destructive commands and gate Edit/Write actions with freeze-boundary checks.
- Specific safety use case: combines warning-before-damage and directory-scoped editing, which gives agents more leverage than a generic prompt.
- Dependency risk: the skill explicitly references sibling /careful and /freeze directories, so it is not standalone.
- Adoption clarity is limited: there are no support files, install command, or extra reference materials to help users verify setup quickly.
Overview of guard skill
What guard does
The guard skill is a full safety mode for Workflow Automation: it combines destructive-command warnings with directory-scoped edit restrictions. In practice, it helps an agent stop before risky actions like rm -rf, force-pushes, or unplanned writes outside the allowed folder.
Who should install it
Install guard if you work near production systems, live data, or tightly bounded code areas and want the agent to be conservative by default. It is especially useful when you want a skill that supports “guard mode,” “maximum safety,” or “lock it down” behavior without relying on a generic prompt alone.
Why it stands out
guard is not just a reminder to be careful. It uses hooks to enforce safety checks at tool-use time, so the constraint is operational rather than purely conversational. That makes it a better fit when you need repeatable guardrails for Workflow Automation, not just a one-off warning.
How to Use guard skill
Install and dependency check
Use the guard install flow for the parent stack, then confirm the supporting skills are present. guard depends on sibling /careful and /freeze skill directories because its hooks call scripts from both. If those are missing, the safety checks will not work as intended.
Give the agent a bounded task
The best guard usage starts with a clear scope, such as: “Update the deployment script only inside services/api/ and warn me before any destructive command.” Include the directory boundary, the type of change, and the level of caution you want. That gives the skill enough context to enforce the right restrictions.
Read these files first
Start with SKILL.md and SKILL.md.tmpl. The first explains the active behavior; the template helps you understand how the generated skill is assembled. Because this repo has no extra rules, references, or helper scripts in the guard folder, these two files are the main source of truth.
Prompt pattern that works
A strong request is specific about both action and safety. For example: “Use guard for Workflow Automation to edit only infra/prod/, warn before any destructive Bash command, and stop if a step would write outside that folder.” This is better than “be careful,” because it defines the exact boundary and expected behavior.
guard skill FAQ
Is guard only for advanced users?
No, but it is most valuable when the cost of mistakes is high. Beginners can use it as a safety wrapper, especially when they are unsure whether a command or edit is risky.
How is guard different from a normal prompt?
A normal prompt can ask for caution, but guard adds enforced checks through hooks. That matters when you want the agent to actually block unsafe actions instead of merely describing them.
When should I not use guard?
Do not use guard if you need broad, exploratory edits across many folders or if the task is intentionally destructive and already reviewed. The directory-scoped behavior can slow down work that does not need strict boundaries.
What should I expect during use?
Expect warnings before destructive commands and checks before edits or writes outside the allowed directory. If your task is outside the configured boundary, the workflow should pause rather than continue blindly.
How to Improve guard skill
Tighten the scope before running
The biggest quality gain comes from naming the exact directory, files, and outcome you want. “Fix the auth migration only in apps/web/src/” produces better results than “make the app safer,” because the skill can enforce the right freeze boundary.
Separate safe changes from risky ones
If your workflow includes both routine edits and dangerous operations, ask the agent to stage them separately. For example, first request read-only inspection and planning, then approve the minimal write set, then handle any destructive action only after confirmation. That makes guard skill behavior easier to verify.
Watch for boundary mismatches
The most common failure mode is an overly broad task description that conflicts with the freeze scope. If the agent keeps stopping, narrow the target path or rewrite the request so the intended files clearly fall inside the protected area.
Iterate with explicit guardrails
After the first run, refine the prompt with the exact commands, folders, or failure conditions you want protected. Good guard usage is iterative: the more concrete your limits and approval rules are, the less guesswork the skill needs and the more reliable the automation becomes.