Baseline Skills: A Security Baseline for AI Coding Agents

Summary
Three open-source skills that give AI coding agents a plain, vendor-neutral security baseline: secure-coding rules, project setup checks and a pre-merge scan, with dependency-free scripts short enough to read.
My part
The problem
AI coding agents write a lot of code quickly, and they repeat the same security mistakes: a secret committed in a config file, a missing access check, a dependency with a known vulnerability. Most teams either have no security baseline at all, or inherit a compliance document too dense to follow. I wanted something an agent actually applies while it works.
The rule it never breaks
No company, cloud or framework baked in, and every script short enough to read before you trust it. The scripts use only Python's standard library, and all three together come to under 500 lines. The checks only read; none of them change your files.
The three skills
- security-baseline. Rules the agent follows whenever it touches auth, secrets, file uploads, APIs, databases, CI/CD or infrastructure, even when nobody says the word "security". It's aligned to OWASP, CIS Benchmarks and NIST CSF / CISA's baseline goals, and written as plain rules rather than a control matrix.
- project-baseline-setup. Scaffolds a new repo with the basics (a README, an
.env.example, a.gitignorethat covers secret files, a lockfile check and a folder for decision records), or audits an existing one and fills only the gaps. It never overwrites a file that already has real content. - pre-merge-check. Before a merge, it scans staged changes for secrets, runs the right dependency audit for the project (npm, pnpm or yarn audit, pip-audit, govulncheck or cargo-audit) and flags debug leftovers and TODOs. Findings are redacted in the output, and it never deletes evidence to make a check pass.
Decisions worth explaining
Prefer the real scanner
The bundled secret scanner is a fallback with ten patterns, not a replacement for a dedicated tool. The skill tells the agent to use a proper scanner when the project has one, and to treat a clean result as "nothing obvious", not a guarantee.
Report, don't guess
If a dependency tool's output can't be parsed, the script shows the raw output instead of inventing a severity. If a tool isn't installed, it says so and shows how to install it, rather than installing anything itself.
Don't act like CI
pre-merge-check reports problems; it doesn't block merges. That's the CI system's job.
Using it
Install with npx skills add Saadaan-Hassan/baseline-skills, or link the skill folders into Claude Code or Codex; each skill installs on its own. It's written in the open Agent Skills format, so it's meant for any agent that supports it. MIT licensed.
Limits
It's an early release. There are no automated tests yet, the secret patterns aren't exhaustive, and severity is only parsed for the npm family of audits; for the other tools, any finding is reported as worth a look. Sharper patterns and stack-specific references are welcome as contributions.
How it was built
I wrote it with Claude Code, and I use the skills in my own setup.

System Visuals

Have a similar engineering problem?
Let’s discuss your requirements, constraints, and what a useful solution would look like.