Skill Vision v1.0.0 — the Claude skill evaluator and scorer

The QA skill for Claude skills: install it, then ask Claude to validate, test, and score your other Agent Skills before they ship.

The launch story: we ran the five validators over a private 20-skill corpus (145 scripts, ~45k LOC). The audit's most valuable output was 13 defects in this harness itself — every critical one is fixed in this release, each with a regression test. A validator that audits itself honestly is the whole point.

What's in v1.0.0

• spec_validator — Agent Skills spec conformance (loads? uploads? triggers?) + per-skill token-cost report (description tokens load every session, body on trigger)
• skill_validator — structure and documentation
• script_tester — executes every bundled script, recursively, with crash-honest verdicts
• quality_scorer — five-dimension score, A+ to F, improvement roadmap
• security_scorer — credentials/injection/traversal posture, false-positive-hardened (anchored patterns, docstring-aware)
• 117 adversarial self-tests, CI on Python 3.9–3.13, zero dependencies (pure stdlib — runs in the no-network Claude API container)

Install

git clone https://github.com/AlveeeRahman/skill-vision.git ~/.claude/skills/skill-vision

Docs: https://alveeerahman.github.io/skill-vision/ · Post your best skill's score in a checkup issue.