Skill Grader
Every installed skill's description rides in the system prompt of every message — whether or not it ever fires — and heavy descriptions dilute the triggering of everything else you have installed. Paste a GitHub repo or a SKILL.md and get a report card on what your skill costs: resident footprint, description honesty, body size, progressive disclosure, factoring, and CLI leverage. Deterministic, free, no signup.
Methodology: “@skills: Attention Is All You Have” (Yin et al., 2026) + the open seo-skill-bench footprint scorer. We graded our own skill first — it got a C+, so we rebuilt it.
FAQ
What does the Skill Grader measure?
Six deterministic axes: resident footprint (the tokens your skill descriptions occupy in the system prompt of every message), description honesty (statement of purpose vs. trigger-keyword padding), body size vs. the published-skill corpus (median 921 words, p90 2,207), progressive disclosure (is detail fetched at point of use?), factoring (one procedure per skill vs. monolith), and CLI leverage (rules pushed into code cost zero attention). Tokens are approximated as chars/4.
Why does an installed skill cost tokens on every message?
A skill’s frontmatter description is resident in the agent’s system prompt so it can auto-trigger — it loads on every message whether or not the skill ever fires. Research ("@skills: Attention Is All You Have", Yin et al. 2026) measured descriptions at 50–280 tokens each and showed heavy descriptions dilute the trigger reliability of everything else installed alongside them.
What grade does SEOAgent’s own skill get?
We graded ourselves first — the original skill scored a C+ (17,237-word body). We rebuilt it against this methodology: the body dropped 70% and every protocol moved into reference files fetched at point of use. Run the grader on @seoagent-official/seoagent and check.