# Prompt Fence

> Scan prompt and instruction files for jailbreak phrases, secret exfiltration cues, and unsafe shell patterns.

- Skill: `sahit-sai/prompt-fence` (Agent Skill, multi-file: 5 files)
- Install (CLI): `npx skillmds@latest add sahit-sai/prompt-fence`
- Raw SKILL.md: https://api.skillmd.com/api/skills/sahit-sai/prompt-fence/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: sahit-sai (https://skillmd.com/u/sahit-sai)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/sahit-sai/prompt-fence

---


# prompt-fence

## Purpose

Catch risky prompt text before it gets bundled into skills, system prompts, or operator playbooks.

## Runbook

1. Scan one file, one directory, or a narrowed skill subtree rather than an entire home directory.
2. Run `scripts/scan.py` to flag common jailbreak phrases, data exfiltration requests, and unsafe shell habits.
3. Review each finding in context; a match is a lead for review, not automatic proof of malicious intent.
4. Escalate only the lines that clearly weaken instruction boundaries or leak secrets.

## Stop conditions

1. Abort if the scan target includes private data that should not be copied into reports.
2. Abort before classifying a file as malicious without manual review.
3. Abort if the current variant cannot safely inspect the relevant local files.

## Output format

- File-by-file finding inventory
- Severity, rule name, line number, and matched excerpt
- A concise markdown or JSON review artifact

## Example invocations

- `python3 skills/prompt-fence/scripts/scan.py skills/`
- `python3 skills/prompt-fence/scripts/scan.py README.md --markdown`