9.70% of Claude Code artefacts fail to load, census finds

A structural census of the Claude Code artefact ecosystem has put a number on its broken files: 43,199 of 445,348 published listings, 9.70%, do not load as published. The figures come from Kynth, which read every skill, subagent, plugin and marketplace it could find on GitHub and skills.sh on 2026-08-14.

At a glance

• The population splits into 347,382 skills, 71,430 subagents, 20,212 plugins and 6,324 marketplaces, and the failure rate varies from 0.16% for plugins to 21.9% for subagents.
• Almost all of the damage is one defect: 38,183 failing listings, 88.4% of the total, carry a YAML frontmatter block that does not parse, so no name or description can be read.
• Nothing in the publishing path catches the defect, since there is no registry gate, no CI convention and no error shown to the author, and every failing row is published under CC BY 4.0.

There has been no prior measurement of this ecosystem, so the 9.70% has nothing to be compared against yet, and its value is as a baseline rather than a verdict. The shape of the finding is the more useful part: the failures appear to concentrate where a runtime enforces required fields, which suggests the gap is a publishing-pipeline problem rather than an authoring one.

Subagents fail at 21.9% and skills at 7.89%

That is 15,669 of 71,430 subagent listings against 27,398 of 347,382 skills. The two manifest-based kinds are near-clean by comparison: plugins fail at 0.16%, 33 listings out of 20,212, and marketplaces at 1.57%, 99 manifests out of 6,324.

Kynth attributes the gap to the frontmatter reference in Anthropic's documentation, which marks a subagent's name and description as Required: Yes while a skill's are not. A skill with no name falls back to its directory name, while a subagent has no identifier to register under. 4,119 failing subagents carry no name and 1,799 no description.

38,183 failing listings carry a YAML block that does not parse

That single defect accounts for 88.4% of the 43,199 failures. Kynth describes it as the cheapest class of bug in software, sitting in front of the most-copied artefacts in a fast-growing ecosystem, because nothing in the publishing path checks it: no registry gate, no CI convention, no error surfaced to the author. Nothing marks the fault, so the file is committed, the repository collects stars, and the artefact never registers.

128 plugin.json or marketplace.json manifests do not parse as JSON, and 99 marketplace manifests parse but list no plugins, so there is nothing to install. A listing can carry more than one fatal cause, which is why the causes sum above 43,199 failures.

The 445,348 listings come from 14,233 repositories

The census counts listings, meaning one artefact in one repository, at an average of 31.3 per repository, because the ecosystem is heavily forked and vendored and the same file is counted once for every repository carrying it. Kynth treats the failure rate as the citable figure and the population count as no claim about how many distinct artefacts exist.

The failures concentrate as well: 3,168 repositories account for all 43,199 of them. The checks are static, nothing is executed and no row is hand-curated, so passing means only that a file parses into something a runtime could register. Discovery is bounded by GitHub's rate-limited search API, and the denominator includes everything found, including what could not be read.

The missing-references flag, corrected

Kynth published a correction alongside the first edition: a missing-references flag fired on 139,623 listings, 40.2% of every skill in the index, and was wrong roughly seven times in eight, from three bugs in one regular expression. Corrected, 18,657 skills, 5.37%, name a genuinely absent bundled file.

The flag is a score deduction rather than a load verdict, so the 9.70% headline is unaffected. The fix sits in the open-source scraper these indexes share and lands at the next nightly rebuild; no date is given for the next census edition.