State of agent skills 2026: what 1,638 public skills look like when scanned

• The median public skill scores 88/100. 46% grade A, 30% grade B and 7% grade F.
• 4% of skills have at least one critical finding, and 19% have a security finding at warning level or above.
• The most common finding is eff/description-trigger, in 34% of skills: the description never says when to use the skill, so the agent has to guess.
• The median skill loads 1,666 tokens when it triggers. 9% load more than 5,000.
• 12% of skills with a body over 40 lines have no numbered workflow.
• 83% come from a repo with a license GitHub can detect.
• 18% bundle at least one script, and 57% link to at least one external URL.
• Only 9% have no findings at all.

75% of public skills score B or better, and 7% get an F. Grades follow the score (A is 90 or more, B 80, C 70, D 60), except that any critical finding makes the grade F whatever the total.

The score is four pillars of 25 points each. Clarity holds up best, at 22.3 on average, and efficiency loses the most points, at 19.6.

The top finding, eff/description-trigger, appears in 34% of public skills. Most of the top ten are cheap to fix: one line in the description, one Output heading, one example.

A skill counts once per rule, however many times the rule fires in it. clarity/workflow only checks bodies over 40 lines, so its share of those skills is 12%. The keys behind spec/non-portable-keys are most often tags, version and argument-hint.

The median public skill loads 1,666 tokens into context when it triggers, and the 90th percentile is 4,708.

The always-loaded cost is small, so installing many skills is cheap until they trigger. The body is the real cost: 9% of skills (145) go past the 5,000 tokens the scanner treats as the budget. Moving long tables and examples into reference files keeps them out of context until needed. See reduce skill token usage and progressive disclosure.

3% of public skills (42) have a critical security finding, and 27% have a security finding of any level.

These are findings, not verdicts. A rule matches a pattern in the text. A skill that documents sudo for an install step, or warns readers about curl | sh, trips the same rule as one that does it. The scanner already downgrades matches that read as a quoted example or a warning to warning level. Each public report shows the file and line so you can judge it in context.

The opposite error also exists. A static scan reads text. It does not run scripts, follow redirects or see what a URL serves later, and a determined author can word around any pattern list. A clean scan lowers risk. It does not prove a skill is safe. Read is this skill safe for what to check by hand, and browse every report on public skills, scanned.

The 67 skills in the ProSkills catalog have a median score of 100, against 88 for public skills. That gap is by design, not luck: catalog skills are scanned by the same rules before they ship and fixed until they pass.

See every catalog score on the quality leaderboard.

The data covers 1,638 skills from 83 public GitHub repositories, scanned on October 19, 2026.

• Selection: GitHub repository search for a seed list of well-known skill repos, the topics claude-skills, agent-skills and claude-code-skills, and repos with "claude skills" in the name or description and over 50 stars. Forks and archived repos are skipped, and the most-starred repos are kept.
• A skill is any folder with a SKILL.md, up to 40 per repo. Folders named test, tests, fixtures or examples are skipped. Up to 25 text files under 200 KB are read per skill.
• When the same SKILL.md appears in several repos, it is counted once, under the most-starred repo. A symlinked SKILL.md is skipped as a copy.
• Each skill is pinned to the commit it was read at, so a report matches a specific version.
• Scoring uses the same scanner as the free Skill Scanner: four pillars of 25 points, findings by named rule, and 42 distinct rules fired in this data. The full rule list is in the validator guide.
• Tokens are estimated as characters divided by four, which lands within about 10% of real tokenizers for English and code.
• License is the repository license as GitHub reports it, not a per-skill license file.

Limits. The sample leans toward popular repos, so it says more about skills people install than about every skill written. The scanner checks form, cost and known risky patterns. It does not judge whether a skill's advice is good. Rule changes rescore every skill at the same commits, so numbers here can move between refreshes without any skill changing.