AUTOMATED EVALUATION · EVIDENCE ATTACHED
SkillGrade automatically evaluates Claude Code skills across security, quality and hygiene — and shows the concrete check verdicts behind every letter, so trust is auditable, not magic.
NEW · MCP SERVER · mcp.skillgrade.dev
Connect the SkillGrade MCP to Claude Code and ask it to check the skills you already have. It matches each SKILL.md against our graded catalog by content hash — so you see the grade, whether your copy drifted from the one we graded, and lightweight notes on what to fix.
Read-only and privacy-first: only a skill's name and a hash ever leave your machine — never the skill's contents. The server runs nothing you send it.
1 · ADD THE SERVER
2 · ASK YOUR AGENT
lookup_skill
One skill by hash or name → grade + findings.
audit_skills
Your whole set at once → summary + per-skill verdicts.
search
Find a graded skill by name.
Status — verified: your copy matches ours · drift: yours differs, we can’t vouch for it · reference: graded, send a hash to confirm · unknown: not in the catalog yet.
HOW WE GRADE
A deterministic pre-check layer gathers facts — file counts, frontmatter, flagged patterns — then a model evaluation grades each dimension against the rubric. The overall grade is simply the worst graded dimension, so one letter can never hide another.
Not evaluated is not a bad grade. A dash means missing data — it is never colored red or green.
F means rejected. Usually a critical security finding; it is information, not punishment.
THE ALGORITHM
Trust you can’t inspect is just branding. So here is exactly how every letter on this site gets made: the same six steps for each skill, whoever wrote it.
Deterministic pre-checks
Before any model runs, code measures the facts: frontmatter validity, file count, size, flagged patterns.
The skill is data, never instructions
Its SKILL.md is evaluated as untrusted content, wrapped so nothing inside it can steer the grader. That’s the same injection-resistance we grade skills for.
Thirty checks, three dimensions
Security (12), Quality (10) and Hygiene (8), from a versioned rubric. A calibrated model rules each check pass, fail, warning or n/a, and points at the evidence.
Scores are computed in code
The model never assigns a letter. Each dimension’s checks aggregate to A–F in code, so a grade can’t be hallucinated or nudged.
Overall is the worst dimension
One letter can never hide another. An F marks a critical finding, usually security. Read it as a pointer to what to check.
Calibrated, vetted graders
We grade with frontier models we keep re-testing (Claude Sonnet and Gemini 2.5 Flash), and disqualify any that false-positive or bend under manipulation. Each skill’s grade names the model that made it.