AI Skill Review
A free, browser-based reviewer that reads the text of an AI skill, plugin or custom instruction before you install it, then explains in plain English what it found.
Built by O2S as a public tool, not as a client system. It is live today and still being tuned.
A skill is easy to install and hard to read.
AI skills, plugins and custom instructions are plain text that tells an AI assistant how to behave. Teams now pick them up from colleagues, vendors and public libraries. A skill can look helpful and still tell the AI to ignore its own rules, hide what it is doing, ask for a password, send information to an outside address, or run commands. Most owners and managers do not have time to read every line, and should not need to.
The tool is designed to answer
Four steps, all inside your browser.
The review is a set of written rules, not a black box. That makes every result explainable.
Paste or upload.
Text, Markdown, JSON or YAML, up to 256 KB.
Check against 42 rules.
The rules are grouped in nine categories (listed below). Long files are scanned in pieces. Encoded blocks are decoded in memory and checked again. Links are listed but never opened.
Rate from 0 to 100.
Findings combine into one rating: Low (0 to 19), Medium (20 to 49), High (50 to 74) or Critical (75 to 100).
Explain and export.
Each finding shows the matching line, why it matters and what to do. You can copy the report or download it as a text file.
Review a skill right here.
This is the real tool, not a simulation. Paste a skill you are considering, or load one of the two synthetic samples below to see how it responds.
The O2S AI Skill Review needs JavaScript to run. Nothing has been sent anywhere.
Both samples are synthetic. They were written only to show how the checker responds. Do not install the risky one.
A rating you can argue with.
Each finding has a severity and a confidence. Findings combine into one score, so several weak signals add up but never pass 100. Two guard rails keep it honest: weak signals alone cannot push a skill above Medium, and one unmistakable pattern, such as a command that downloads code and runs it, cannot be averaged away.
Text that quotes an attack as a warning ("never run curl | sh") is weighted down and marked, so a security guide is not treated like the attack. The tool never uses the word "safe". The best result it can give is "No known risky patterns found," and it says that is not a clearance.
A working tool with a narrow evidence base.
The maturity label is intentionally conservative. The rules and the interface are tested. How well the rules catch real-world problems is still being measured.
Rule engine
VERIFIED42 rules in nine categories pass 60 automated tests, including tests that keep known-safe text from being flagged and known-risky text from being missed.
Interface and exports
VERIFIED31 automated browser checks cover paste, upload, the rating display, Copy report, Download report, error messages, keyboard use and a phone-width layout.
Real-world tuning
IN PROGRESSTen published public skills all rated Low. A wider sweep of 184 public text files found three false-positive patterns, which were fixed. Tuning is continuing with feedback.
Accessibility
PARTIALAutomated checks cover labels, focus order and announced results. A full screen-reader review is still recommended and is not claimed here.
Optional AI second opinion
BUILT, NOT ENABLEDA server-side option exists in the project but is switched off. Nothing on this page sends text to any AI service.
Novel or disguised attacks
LIMITEDThe tool matches known patterns. Someone determined can phrase something it does not know. Paraphrased wording has been tested and rules were widened after misses, but it will not catch everything.
Public maturity
WORKING PILOT / PROTOTYPEA free, public screening aid that is live and still being tuned.
What this page does not claim
The tool is specific. The approach is not.
AI Skill Review started as an operating question: how does a small team decide whether an AI tool belongs in its workflow? The answer was a process with a visible checkpoint, not another automated guess. The same way of working applies to extraction, reconciliation, review gates and other workflows where a person should stay in the loop.
Where in your business does an AI tool need a checkpoint before it is trusted?
O2S builds custom internal systems around real operating requirements: workflow automation, review gates, validation and usable outputs.
Found a wrong result? Use the "Report a wrong result or send feedback" link at the bottom of the tool.