Skilltest — Testing skill for Claude Code

SkillTest

**Test framework and CI gate for AI agent skills (`SKILL.md`).** Bilingual RU/EN evals · behaviour contract (`spec.yaml`) · LLM judge · with-skill vs no-skill baseline · trigger test · prompt-injection guard with Cyrillic-aware patterns · HTML / JSON / JUnit / SARIF reports · non-zero exit codes for CI.

[](https://github.com/archplg/skilltest/actions/workflows/ci.yml) [](LICENSE)

**→ [skillemall.ai](https://skillemall.ai)** is the public rating this engine runs: 118,000+ skills pulled from open catalogs, graded A–F for safety and quality, with a second, separate grade for whether a skill's process actually runs to the end. Search it, browse today's catch of broken or risky skills, or check your own skill there for free — no install. Everything below is the same engine, to run yourself or wire into CI.

[Русская версия → README.ru.md](README.ru.md)

Until the package is on npm: `npm i -g github:archplg/skilltest`, then `skilleval …`; or `npx github:archplg/skilltest …`.

npx skilleval init  ./my-skill      # scaffold spec.yaml + evals/evals.json
npx skilleval run   ./my-skill      # guard → cases → triggers → report (exit 0/1/2/3)

npm package: **`skilleval`** (the name `skilltest` was taken). Both `skilleval` and `skilltest` binaries are installed.

Why

Skills are text. When you edit one, nothing tells you it stopped working: no stack trace, no red status, just a behaviour change somebod