Skill Grader
Free grader for Claude Code / agent skills — report card on token footprint, honesty and factoring.
Last verified:
What is Skill Grader?
Skill Grader is a free tool that analyzes skill definitions from GitHub repos or SKILL.md files and grades them on six deterministic axes: resident footprint, description honesty, body size, progressive disclosure, factoring, and CLI leverage. It helps identify how much token overhead each skill adds to system prompts and suggests optimizations to reduce bloat without signup.
Skill Grader pricing
Pricing model: Freemium
Free, no signup
Skill Grader pros
- Completely free with no signup required
- Deterministic scoring based on published research and open benchmarks
- Identifies concrete optimization opportunities (footprint, disclosure, factoring, CLI leverage)
- Backed by the open seo-skill-bench footprint scorer and peer research (Yin et al., 2026)
- Actionable metrics—creators rebuilt their own skill after initial C+ grade and cut body size by 70%
Skill Grader cons
- Limited to SKILL.md format; may not work with other skill definition systems
- Measures structural and descriptive qualities only, not functional correctness or actual utility
- Grading methodology is opinionated (resident footprint minimization is not always optimal for discoverability)
Frequently asked questions about Skill Grader
What does the Skill Grader measure?
Six axes: resident footprint (tokens in system prompt), description honesty, body size vs. published corpus (median 921 words, p90 2,207), progressive disclosure, factoring (one procedure per skill), and CLI leverage.
Why does an installed skill cost tokens on every message?
Skill descriptions are resident in the agent's system prompt for auto-triggering, loading on every message regardless of whether the skill fires. Research shows descriptions cost 50–280 tokens each and dilute trigger reliability of other skills.
What grade did SEOAgent's own skill get?
Original: C+ (17,237-word body). After rebuilding with this methodology: body dropped 70%, protocols moved to reference files fetched at point of use.