Gaia Skill Tree Meta Audit Series

This report covers three user-facing changes to the skill tree this cycle. The most significant is a registry correction: mattpocock/caveman was removed after its author deleted it upstream and confirmed it was never meant to be public, and the skill it was a private copy of — juliusbrussee/caveman, the real original — was ingested in its place at a conservative, evidence-checked 1★. Two skills also picked up clearer install slugs that match what their upstream repositories actually ship. And a batch of gaia skills install targets that used to silently report success while installing nothing now correctly report that they can't be installed.

caveman: a registry correction, not just a removal

mattpocock/caveman sat in the registry at 2★. Its upstream repository deleted the skill on 2026-05-31, and its author confirmed directly that it was a private duplicate of another skill he was testing — never meant to be public, with no replacement. That much was a straightforward removal.

The more interesting part is what it was a duplicate of. Tracing the skill's actual history turned up JuliusBrussee/caveman — MIT licensed, over 96,000 stars, actively maintained, homepage at caveman.so. The evidence that this is the true original, not the other way around:

• Julius Brussee's caveman has a first commit dated 2026-04-04. The copy that appeared in

mattpocock/skills doesn't show up until 2026-04-17 — thirteen days later.

• Both SKILL.md files share the same "~65-75% token reduction" framing, the same rule

structure, and at least one verbatim example line.

• When mattpocock/skills deprecates a skill for a real reason, its changeset names a

replacement — it did exactly that for four other removed skills around the same time. Caveman's removal named none, consistent with it being a private test copy rather than original work with a designated successor.

juliusbrussee/caveman is now the registry's caveman entry, and the honest story is worth telling in full because it's not simply "verified, ship it." The skill's own marketing claims a ~65-75% reduction in agent context/token usage. Independent measurement tells a different story: one measured source puts the real reduction closer to 31%, and another closer to 9%. That gap between the claim and the independently measured result is exactly why the skill landed at a conservative 1★ rather than higher — a real, popular, well-maintained skill, whose central performance claim doesn't hold up to outside scrutiny at the strength advertised. Readers evaluating whether to install it should weigh the skill's genuine popularity and maintenance against that specific, checkable gap.

caveman is installable from JuliusBrussee/caveman via the usual gaia skills install juliusbrussee/caveman.

Two skills got install-accurate slugs

Two named skills were renamed so the slug you install by matches the name the upstream project actually ships under:

browserbase/stagehand had been flagged as possibly having no installable skill manifest at all. That turned out to be wrong — the manifest exists upstream, just under a different skill name (browse), and the registry now points at it correctly. bradautomates/watch picks up the name its own SKILL.md uses rather than the older working title. Anyone with either skill already installed keeps working; anyone installing fresh should use the new slug.

Install failures that used to lie now tell the truth

A number of named skills in the registry pointed links.github at a location that doesn't actually contain an installable skill — an issue thread, a general framework repository with no skill manifest, or a directory that never had one. Previously, running gaia skills install against one of these could report success while installing nothing usable. That's now fixed: these skills are correctly marked as not installable, so gaia skills install gives you an honest failure instead of a false "done."

This is a trust fix, not a data-tidying exercise — the old behavior meant you could believe you'd installed a skill you didn't actually have. A related cleanup pass is still in progress for a further set of skills where the right call needs a maintainer's judgment call (for example, whether a skill grouped under a component suite should be held to the same individual-installability bar as a standalone skill); those are being worked through deliberately rather than rushed.

Reading the data

The key facts are:

• mattpocock/caveman removed; juliusbrussee/caveman — the verified true original —

ingested at 1★, with the token-reduction discrepancy disclosed rather than smoothed over.

• browserbase/browse (was stagehand) and bradautomates/watch (was claude-video) are

the current install slugs for those two skills.

• Several previously silently-broken install targets now fail honestly instead of falsely

succeeding.

• The registry currently tracks 180 named skills across 38 contributors.

Sources and References

[1] JuliusBrussee/caveman — https://github.com/JuliusBrussee/caveman — the verified true original source now backing the registry's caveman entry.

[2] mattpocock/skills deprecation commit — https://github.com/mattpocock/skills/commit/7d3ada9716a9ee08d6c6f775d8a78ef889e1798f — upstream deletion of the duplicate copy, authored by Matt Pocock.

[3] Browserbase — https://github.com/browserbase — upstream source for the browse skill now installed as browserbase/browse.

[4] Brad Automates — https://github.com/bradautomates — upstream source for the watch skill now installed as bradautomates/watch.