Inspect before you delegate
Agent Skill Reviewer
Read the fine print in your agent's instructions. Find risky patterns in a skill file and see exactly where to look next.
Your skill instructions
Local onlyPaste SKILL.md, agent rules, or other instructions. The text stays in this tab; it is never executed, uploaded, or saved by the reviewer. Limits: 100 KB per file and 4,000 characters per logical line.
Review findings
What is this skill asking your agent to do?
Run a review to trace sensitive access, execution, and permission patterns to specific lines.
15 explainable checks. No account needed.Know the coverage
A first pass through the instructions.
A skill can reference other scripts, dependencies, and services. Review those separately and verify the permissions the runtime actually grants. A clean result here is only the beginning of that review.
Pattern matching focuses on English instructions and common command syntax. Context labels help interpretation; they never suppress a match or reduce its priority.
See all 15 checks Rules v1.0.0
- Downloaded code passed to an interpreter
- Instructions try to override existing rules
- Approval or permission checks bypassed
- Credential files or secret stores referenced
- Possible disclosure of sensitive data
- Elevated privileges or host access
- Destructive command pattern
- Integrity or transport verification weakened
- Encoded or dynamic execution
- Changes may persist beyond the task
- Actions or evidence hidden from the owner
- Loop or spending limits are unclear
- Broad tool access declared
- A moving package version is requested
- Invisible or directional control characters
Using the reviewer
What does an Agent Skill Reviewer check?
It scans a supplied instruction file for visible patterns involving downloaded code execution, secret access, approval bypasses, destructive commands, persistent changes, and instruction overrides. Every match includes a line reference and a suggested action.
Does my SKILL.md leave the browser?
The reviewer does not upload, save, or execute your input. Local files are read in this tab. If you consent to site analytics, tool events contain completion and finding counts, never source text or filenames.
Can this prove that a skill is safe?
No. This is a static pattern checker, not a model evaluation or malware sandbox. It can miss unfamiliar, encoded, multilingual, or indirect instructions and can flag legitimate examples. Referenced scripts, dependencies, links, and actual runtime permissions need separate review.
Can I review instructions from different coding agents?
Yes. Paste Markdown or plain-text instructions from your agent workflow. The checks are text-based and do not validate product-specific configuration syntax or guarantee that an agent will enforce an instruction.
Further reading
Use these primary references for the wider review: Agent Skills file structure, OWASP agent security, and OWASP prompt injection prevention. These organizations do not endorse this tool.