# Guardrail Policy Writer

> A Claude skill from 45ck/llm-agent-security-skills.

- **Type:** Skill
- **Install:** `agentstack add skill-45ck-llm-agent-security-skills-guardrail-policy-writer`
- **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior
- **Seller:** [45ck](https://agentstack.voostack.com/s/45ck)
- **Installs:** 0
- **Category:** [AI & ML](https://agentstack.voostack.com/c/ai-and-ml)
- **Latest version:** 0.1.0
- **License:** MIT
- **Upstream author:** [45ck](https://github.com/45ck)
- **Source:** https://github.com/45ck/llm-agent-security-skills/tree/main/skills/guardrail-policy-writer

## Install

```sh
agentstack add skill-45ck-llm-agent-security-skills-guardrail-policy-writer
```

Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent.

## About

# guardrail-policy-writer

## Purpose
Write explicit agent guardrail policies covering allowed actions, prohibited behavior, escalation points, approvals, and safe fallbacks.

## Trigger this skill when
- The current artifact has LLM, prompt, retrieval, memory, or agent-tooling behavior that needs structured review or hardening.
- You need to turn vague AI safety or agent security concerns into concrete findings, controls, or requirements.
- You want the next agent-security-focused action to be explicit rather than ad hoc.

## Expected inputs
- agent role
- risk appetite
- tool permissions
- failure modes
- compliance or business constraints

## Deliverables
- guardrail policy
- allowed/prohibited action tables
- escalation rules
- fallback behavior
- recommended next skill

## Operating procedure
1. Read the artifact from an agent-security perspective and identify the concrete trust, autonomy, and side-effect model.
2. Separate facts from assumptions and call out missing details that materially affect risk or confidence.
3. Start with the highest-impact abuse paths rather than trying to describe every possible issue equally.
4. Translate findings into explicit controls, approvals, isolation boundaries, or policy language that builders can act on.
5. Prefer concrete exploit paths, sinks, boundaries, and failure conditions over generic AI-safety slogans.
6. Finish with the most sensible handoff skill based on the dominant risk pattern you found.

## Quality gates
- Findings are specific to the actual prompt, retrieval, memory, tool, or runtime design rather than generic AI risk boilerplate.
- Output separates facts, assumptions, risks, controls, and recommended next action.
- Prioritization reflects impact, privilege, and automation potential rather than just issue count.
- Recommendations are implementable and framed in a way that can be tested or reviewed later.

## Handoff targets
- tool-permission-boundary-checker
- agent-agency-reviewer
- agent-sandbox-reviewer

## Output style
- Be explicit about uncertainty.
- Prefer concrete abuse paths and control implications over generic safety slogans.
- Separate facts, risks, recommendations, and next steps.
- Make the output usable by engineers, reviewers, security testers, and policy owners.

## Failure modes to avoid
- Do not treat "the model should know better" as a security control.
- Do not bury high-impact autonomous action risk behind long, unprioritized issue lists.
- Do not recommend guardrails without explaining the exploit or failure they address.
- Do not hide uncertainty when prompt assembly, retrieval, memory, or runtime details are missing.

## Minimum output skeleton
```md
## Summary
## Findings
## Structured outputs
## Risks
## Recommendations
## Recommended next skill
```

## Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [45ck](https://github.com/45ck)
- **Source:** [45ck/llm-agent-security-skills](https://github.com/45ck/llm-agent-security-skills)
- **License:** MIT

Install and usage instructions live in the source repository linked above.

## Pricing

- **Free** — Free

## Security capabilities

Automated source analysis of v0.1.0 — what this tool can access:

- **Network access:** no
- **Filesystem access:** no
- **Shell / process execution:** no
- **Environment & secrets:** no
- **Dynamic code execution:** no

*"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.*


## Versions

- **0.1.0** — security scan: passed — Imported from the upstream source.

## Links

- Listing page: https://agentstack.voostack.com/l/skill-45ck-llm-agent-security-skills-guardrail-policy-writer
- Seller: https://agentstack.voostack.com/s/45ck
- Browse the marketplace: https://agentstack.voostack.com/browse

---
Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.