Add claude-code-security-baseline skill

[Miko Builds (mikobuilds)]

What it does

claude-code-security-baseline sets up the security baseline for a project before an agent starts working in it: a permissions.deny list in .claude/settings.json, a gitleaks pre-commit hook, an .env.example that documents a read-only database role, prompt-injection rules appended to CLAUDE.md, and .env added to .gitignore.

The problem it solves

Letting an agent into a repo is fine until it isn't - git reset --hard over uncommitted work, a key that lands in a public repo, a destructive query against production because a stale migration looked wrong to it. None of that takes a clever attack. It takes one confident agent and one tired developer clicking "yes" at two in the morning. Everyone means to set those four files up, nobody remembers the exact syntax, so it gets skipped.

Who uses it

Solo devs and small teams starting a repo, or pointing Claude Code at an existing one. The checklist behind it came out of my own projects and was published first; this skill is that checklist automated.

Safety

Nothing is ever overwritten. An existing settings.json keeps its own keys and gains only the deny rules it's missing, an existing CLAUDE.md gets a section appended, and a second run reports "left alone" and changes nothing. --check writes nothing at all and only reports what's missing. Standard library Python, no network calls, no telemetry - which matters, given what it's touching.

Example

$ python ~/.claude/skills/claude-code-security-baseline/scripts/harden.py --check

[ ] .claude/settings.json    no deny list - nothing stops rm -rf or a read of .env
[ ] .pre-commit-config.yaml  no secret scanner - a forgotten key would reach git
[i] .env.example             already there, left alone
[ ] CLAUDE.md                no security section
[ ] .gitignore               missing: .env.local, .env.*.local

4 of 5 items need attention.

It also prints the two traps that make the whole setup silently useless: edits to .claude/settings.json only take effect from the next session, and a pre-commit config scans nothing until pre-commit install has actually been run. Related trap, since people test it this way: don't check gitleaks with AKIAIOSFODNN7EXAMPLE, it's on the allow-list and sails straight through.

Where I put it

Security & Systems, first alphabetically. Linked to the source repo rather than vendored here, same as the other entries in that section.

Repo: https://github.com/mikobuilds/claude-code-security-checklist (MIT)