anthropics/skills, tested in 1 task class

Skills from anthropics/skills were tested in 1 kind of task, and 1 more has a skill picked and not yet run. One kind of task at a time, and never as one number.

Across the classes it was tested in

How many task classes each result fell in. Counts of classes, not scores.
What happenedClassesWhich
Helped on some1Skill authoring

We do not add these together. A skill-authoring score is a share of requirements met. A voice score is how often a blind reader preferred it. They are different measurements of different things, so a single number over both would be a figure neither test produced. The numbers are on the skill pages, where each was measured.

Per model

Claude Haiku 4.5

Claude Haiku 4.5: classes by result.
What happenedClassesWhich
Helped clearly1Skill authoring

GPT-5 mini

GPT-5 mini: classes by result.
What happenedClassesWhich
Helped clearly1Skill authoring

Gemini 3.1 Flash Lite

Gemini 3.1 Flash Lite: classes by result.
What happenedClassesWhich
Helped a little, under our bar1Skill authoring

Each model is named by the string the vendor actually served, not by the one the run requested.

Coverage, class by class

What was done in each class for this repository. The full grid is on the coverage page.
ClassStateWhat that means
Skill authoringTestedWe ran this one. The same tasks were done with the skill loaded and again without it, and the numbers are on that skill's own page.
Accessibility auditNo qualifying skillWe looked through this repository for this kind of task and found nothing that met the bar we publish.
Spec writingPinned, not runWe found a skill here worth testing and fixed the exact version to test, and we have not run it yet.
On-page auditNo qualifying skillWe looked through this repository for this kind of task and found nothing that met the bar we publish.
VoiceNo qualifying skillWe looked through this repository for this kind of task and found nothing that met the bar we publish.
Code reviewNo qualifying skillWe looked through this repository for this kind of task and found nothing that met the bar we publish.
Test writingDeclared emptyWe said in advance that we would leave this square empty, so nothing here has been weighed against the bar.

Identity pins

Every skill from this repository that holds a slot, and the commit its arm was given.
ClassPathCommitContent hashDate tested
Skill authoringskills/skill-creator3b3fad96af16e51ce07ea7fd2026-08-30

Licence status, per pin

What the licence check found beside each set of pinned bytes, run against that pin's own commit.
Pinned bytesLicenceLicence fileChecked at
skills/skill-creator slotApache-2.0skills/skill-creator/LICENSE.txt3b3fad96af16 2026-08-31
skills/doc-coauthoring rejectedNo licence filenone3b3fad96af16 2026-08-31
skills/brand-guidelines rejectedApache-2.0skills/brand-guidelines/LICENSE.txt3b3fad96af16 2026-08-31

We record the licence for each file we tested, not for the whole repository. Some repositories have no licence at the top and license each skill folder on its own, and one of the files here has no licence at all. Each row is a check against that file's own commit. A row reading not recorded means we have not run the check on those bytes yet. A row reading no licence file means we ran it and found none. Nothing here is guessed from a repository name.

  • skills/skill-creator: Apache-2.0, in a LICENSE.txt inside the skill directory. anthropics/skills carries no licence at its repository root, so this per-skill file is the only licence covering these bytes. The file is byte-identical to the one beside skills/brand-guidelines, blob 4f881c52d1f72f4cfb720e339e2d35c3058d01a9.
  • skills/doc-coauthoring: license: none found; bytes retained in this private repository for measurement only, never rendered
  • skills/brand-guidelines: Apache-2.0, in a LICENSE.txt inside the skill directory. anthropics/skills carries no licence at its repository root, so this per-skill file is the only licence covering these bytes. The file is byte-identical to the one beside skills/skill-creator, blob 4f881c52d1f72f4cfb720e339e2d35c3058d01a9.

Everything on this page is worked out at build time from the runs we committed. We publish only the file we tested and where it came from. No description, no README, no star count, no install line.