CheckSkills @CheckSkillsDev

Last, look for text you can't see. Invisible Unicode and instructions buried in HTML comments can steer an agent and still slip past a quick skim. Or paste the GitHub link into CheckSkills and compare. It's an automated scan, not a guarantee. checkskills.dev/check/?utm_sou…

Check hooks and allowed-tools. Open .mcp.json if it has one. Look for sudo, chmod 777, keys or tokens written into files, and any line that tells the agent to skip permissions or ignore safety rules.

How to read a skill before you install it. A short checklist, built from what the CheckSkills scan looks for. Open the whole folder, not just SKILL.md. Your agent can read every file in there, so you should too.

What to check before you install a skill like this. Is there a .env or any file with a key in it? A key in a public repo is leaked. Don't use it. Get your own. Does it ship node_modules? That's a lot of other people's code you'd run without reading.

Why a C and not an F? An F is for a skill that reads keys, .env or .ssh and sends data to an outside URL. The scan didn't flag that here. This skill ships its own .env, and its code sends that key to the music API it's built for.

One CheckSkills scan, walked through. The skill makes music through an outside music API. Its folder holds a SKILL.md, index.js, package.json, package-lock.json, a node_modules folder, and a .env file.

@houndsecurityai Good call. We're testing a scan that covers every file in the skill folder, not just SKILL.md. It's not live yet. Until then, pointing people to your site as an extra check makes sense. CheckSkills is free, and it'll stay free.

A skill can tell your agent to turn off its own permission prompts. The clawd skill in modbender/skill-library-mcp does it 4 times. Line 7 of a file in its folder says to always start Claude Code with this: claude --dangerously-skip-permissions exec "..."

How to spot it. Search the whole skill folder for --dangerously-skip-permissions and bypassPermissions. Here it sat in a side file next to SKILL.md, not in SKILL.md itself. CheckSkills grade D, last scanned 2026-10-17. checkskills.dev/skills/modbend…

That flag tells Claude Code to stop asking before it runs commands or edits files. Anthropic's docs say to use it inside a container or VM. Inside a skill, it means the agent runs commands without asking you. That includes commands it picked up from text in a file it read.

@CobaltHere @BentoBoiNFT Fair. If you do grab one from GitHub, paste the link here first. It reads the skill file and flags things like a remote script piped into a shell. checkskills.dev/check/?utm_sou…

@hobbessz @iamlukethedev Can't speak for the whole hub. If a skill is on GitHub, paste its link here before you install. It reads the file and grades it A to F with the reasons. checkskills.dev/check/?utm_sou…

@Ken_LinkWell Good list. I'd add one step and paste the skill's GitHub link here. It flags scripts piped into a shell and rules telling the agent not to ask you. checkskills.dev/check/?utm_sou…

@themuscleai That's a lot of SKILL.md files to trust on stars alone. You can paste any of those repo links here and see each skill's grade and why it got it. checkskills.dev/check/?utm_sou…

@stretchcloud Agreed, stars tell you nothing about what a skill runs. You can paste a skill repo link here and get an A to F grade per skill file, with the reason. checkskills.dev/check/?utm_sou…

@Voxyz_ai Good habit. For a quick first pass, paste the repo link here. It grades each SKILL.md A to F and says why. It's automated, so your read still matters. checkskills.dev/check/?utm_sou…