Claude Skill Suite — Security skill for Claude Code

claude-skill-suite

A suite of [Claude Code](https://claude.com/claude-code) skills built to keep a **cheaper model** doing high-quality engineering work. Some are deterministic Node tools — the model just runs them, so quality doesn't degrade with the model; others are prose skills that shape how the model reviews, advises, and writes. The governing rule for the security ones: **skill the judgment, hook the gates** — a risk that must never slip through (a committed secret) belongs in a deterministic hook that fires regardless of which model is running, not in a skill the model has to remember to invoke.

Memory / documentation-system skills live in the sibling repo **[claude-project-memory](https://github.com/Evan-Daruwalla/claude-project-memory)**. Together these two repos hold the generalized skill set.

Security gates — deterministic, hook-backed

Skill	What it does
history-leak-scan	Scans a repo's full git history (or the staged diff) for leaked credentials — per-provider keys (AWS, GitHub, Slack, Google, Stripe, Alpaca), private-key blocks, JWTs, high-entropy assignments, weak passwords. Token-level placeholder suppression so it doesn't cry wolf. Ships a --canary self-test.
commit-gate	Blocks any commit that stages a secret, two ways: a native git pre-commit hook (shell commits) and a PreToolUse hook (commits the model makes via Bash) — because a PreToolUse hook alone misses shell commits. Both share the scanner and fail open on error.

Model quality & output

Skill	What it does
llm-eval-harness	Measures how far a cheaper model falls from your flagship model's quality bar, deterministically: checks (format / no-fabrication / surgical-scope assertions — no golden needed) and golden (word-level similarity to a captured flagship reference). Appends every run to a ratchet so the gap is a tracked series. No LLM-judge — a non-reproducible judge would be invented data.
**token