In plain English
The assignment did not need an animation plugin. That plugin pulled in a second package. The second package calls a raw IP address and runs whatever comes back. Following the README is enough.
Interview assignments often arrive as a full GitHub repo. The README
usually says to run npm install. That can run hidden code
on your laptop.
RepoSafety is a Cursor command. It reads the repository first — without installing or running it — and tells you whether to install. You still decide. A clean report is not a promise that the repo is safe.
/interview-repo-safety
You get a take-home. You clone it. You install packages. You start the app. Maybe you ask an AI coding assistant to work in the folder. Any of those steps can run code you have not read.
RepoSafety sits after clone, before install.
Copy the take-home to your machine. Do not run
npm install, yarn, pnpm,
pip, or the README’s “getting started” commands yet.
Open the folder and type /interview-repo-safety in
Agent chat. The skill only reads files. It does not install the
project to inspect it.
You get one answer: safe to install, review first, or do not install — plus a short report. You choose what happens next.
Every row was checked against the public
/interview-repo-safety skill. A tick means the skill
performs that check. A cross means it does not — even if an older
version of this page implied it.
Looks at files only. No npm install, no tests, no
app start, no Docker of the take-home itself.
Reads package.json and lock files. Also looks at
Python, Rust, Go, and other manifest files when they are there.
Packages your packages depend on are included when a lock file
can be read.
Flags preinstall, postinstall,
prepare, and similar keys — plus app scripts that
would load a trap on npm run dev.
Looks at .git/hooks, VS Code / Cursor tasks, IntelliJ
configs, and Dev Container commands that run when you open the
folder.
Checks .npmrc, .env files, CI config,
and git history. The report names the type, file, and severity.
It does not print the secret.
Static search for things like eval,
new Function, hidden loaders, and hardcoded IPs.
It records the addresses. It does not download the payload.
If a wallet take-home ships a random animation plugin, that mismatch is a finding — not automatic proof of malware, but a reason to stop.
Simple name checks (think lodash vs
lodahs). This is a heuristic, not a full commercial
database.
Exactly one of: safe to install, review first, or do not install — plus high / medium / low confidence. No made-up “safety score.”
After a “safe” or “review” verdict, you can request a sandbox
run. The box sees only the repo folder. The internet is blocked
except registry.npmjs.org until you name another
host. A “do not install” verdict refuses this unless you
clearly ask to observe.
The skill says this outright. A clean report means “nothing obvious turned up,” not “this is safe.”
Those tools are used only if you already have them in Cursor. The skill does not install them. The published example ran without them.
The skill has its own checks. It is not a new dashboard for Sonatype, Opsera, or npm audit.
Heavily obfuscated code and behavior that only appears after install can be missed. File reading cannot see runtime.
The sandbox logs which hosts were contacted and how many bytes moved. It does not decrypt HTTPS.
If the project needs your home folder, sudo, Docker-on-Docker, or an open network, the skill will say you need a disposable VM. It will not loosen the sandbox.
This site explains the skill. The check runs in your editor, on a repo you already cloned.
A coding-interview repo presented as “Proof of Dev.” The README said
to run npm install && npm run dev. That was enough to
trigger the trap. The skill said do not install
before anything ran.
The assignment did not need an animation plugin. That plugin pulled in a second package. The second package calls a raw IP address and runs whatever comes back. Following the README is enough.
Treat that machine as compromised. The skill cannot undo an install that already happened. Do not run this project in the sandbox either — the verdict was do not install.
This is a Cursor skill (it also works in Claude Code). It is not a website you paste a GitHub link into.
git clone https://github.com/naval200/interview-repo-safety.git
cd interview-repo-safety
mkdir -p ~/.cursor/skills
ln -s "$(pwd)/skills/interview-repo-safety" ~/.cursor/skills/interview-repo-safety
So Cursor picks up the new skill.
Open that folder in Cursor.
/interview-repo-safetyRead the verdict before you follow the README.
Docker is only needed later, if you ask to install inside the sandbox.
Other agents can use
npx skills add naval200/interview-repo-safety.
/interview-repo-safety — then you decide.