Open editor
Skill inspector

Inspect agent skills before you install them.

Paste a GitHub link to one skill or a whole collection. See the spec check, the estimated context cost and the scripts worth reading, in a few seconds.

Or open a zip, a folder or the built-in example. Nothing is uploaded and no script is ever run. To fix what the report finds, use “Edit a copy”: it opens the files in the editor and shows a diff against the original.

A real report

What you get for a whole collection.

This is the report for a public repository of 20 skills, read in about seven seconds. Installing all of them puts about 2.4k tokens in every session before you type a word.

github.com/anthropics/skills, 20 skills, run on 13 November 2026. Token counts are estimates.
SkillWhat it says it doesAlwaysOn triggerNotes
claude-apiReference for the Claude API / Anthropic SDK: model ids, pricing, parameters…~270~26k1 REVIEW
webapp-testingToolkit for interacting with and testing local web applications…~55~9701 REVIEW
xlsxUse this skill any time a spreadsheet file is the primary input or output…~238~2.1knone
pptxUse this skill any time a .pptx or .potx file is involved in any way…~184~5.2k1 note

The two REVIEW items: claude-api has a description of 1068 characters, while the specification allows 1024. webapp-testing has a script, scripts/with_server.py, that starts a command through the shell (shell=True), with a comment in the file explaining why. Both are pointers to read, not verdicts. The note on pptx is that its SKILL.md is above the specification’s suggested size.

What it checks

Four things, with fixed rules.

The specification

The frontmatter against the published specification: the name (lowercase letters, digits and hyphens, up to 64 characters, matching the folder), the description (required, up to 1024 characters) and the optional fields. Notes for a SKILL.md over 500 lines or about 5000 tokens, which the specification advises against.

The context cost

Estimated tokens at three levels: the name and description every agent reads at startup, the whole SKILL.md once the skill triggers, and every text file. Estimated from character count, about four characters per token.

Missing files

Files that the instructions mention but the folder does not contain: paths with a folder, link targets, and script names after a command such as python. Example names like your_script.py are ignored.

Scripts worth reading

Scripts matched against a fixed list of patterns: running a command through a shell, running text as code, piping a download into a shell, deleting broad paths, naming places where keys are kept, decoding then running text, long encoded blobs. Network use and package installs are noted, not flagged.

Limits

What it does not do.

It does not run anything, and it is not a malware scanner. A clean report does not make a script safe, and a flag does not make one harmful: the rules are patterns, so they miss things and flag harmless ones. It does not check that a skill does what its description says. Token counts are estimates, and real counts depend on the model. Read what it points at.

Questions

About the inspector

What exactly does it check?

The SKILL.md frontmatter against the published specification, the estimated token cost at three levels, files that the instructions mention but the folder does not contain, and scripts matched against a fixed list of risky patterns.

Does it run the scripts?

Never. It only reads text.

Does it work with private repositories?

Not from a link. Download the repository as a zip and open the zip in the inspector instead.

Can I fix a skill here?

Yes. “Edit a copy” opens every file in the editor, with the checks beside it and a diff against the original. Download your changes as a zip. mdprev never writes back to GitHub.

Where do the files go?

Nowhere. For a GitHub link, your browser fetches the files straight from GitHub. For a zip or a folder, they are read locally. Nothing is sent to mdprev.

More: What is inside a SKILL.md · How much context does a skill cost?