The 19 folders share a format while their runtimes vary
The 430-file checkout at commit 8a1541c contains 19 top-level skill folders. Each starts with SKILL.md, where a name and description tell Claude when to load the instructions. Some stop at guidance. Others include Python or JavaScript, templates, schemas, fonts, and reference files. That makes the repository useful for learning the format, but it does not give the catalog one shared runtime or one setup command.
At about 19,701 lines of source, the checkout occupied 11.7 MB. The range is wide: there are compact instructions for brand or communication work, plus larger document skills for DOCX, PDF, PPTX, and XLSX files. The repository also includes a basic template and the Agent Skills specification. Read the folder you want, because the root README cannot describe every program a skill may run.
Claude Code installs 2 bundles before task-specific setup begins
The 430-file repository's README gives Claude Code 2 install choices: document-skills and example-skills. You first add anthropics/skills as a marketplace, then install the bundle you want. Paid Claude.ai plans already expose the examples, while API users follow the Skills API route. Those are convenient entry points, but installation only makes the instructions available. It does not prove that every local dependency or output path works on your machine.
The measured GIF skill shows the second layer. Its requirements file names 4 packages: Pillow, imageio, imageio-ffmpeg, and NumPy. The prose install snippet lists only 3 and omits imageio-ffmpeg. No Slack token is needed because the code writes a GIF rather than uploading one. Another folder may need office software, browser control, input documents, or service credentials, so permissions should be reviewed skill by skill.
What happened when we ran it
We measured the Python project in skills/slack-gif-creator at commit 8a1541c. Installing 39 packages took 23 seconds and left 190 MB on disk. The build completed in 13 seconds. pip-audit reported 0 known vulnerabilities. Our measurement setup was a fresh Python 3.12 Debian container with 3 CPUs, 8 GB of RAM, no secrets, and no elevated privileges.
The test step in our run failed with exit 5 after 8 seconds. Pytest reported 0 passed and 0 failed out of 0, then ended with no tests ran in 0.00s. That is neither a broken assertion nor a passing suite. The repository contained a tests directory, but the invoked command collected no cases. We cannot turn an empty run into evidence that the GIF helpers or the other 18 skills behave correctly.
The GIF helper checks dimensions but leaves output judgment to you
The 4-package Slack GIF skill provides a builder, easing functions, frame helpers, and a validator. Its documented target is 128 by 128 pixels for emoji and 480 by 480 for message GIFs, with 10 to 30 frames per second. The builder can reduce colors, remove near-duplicate frames, and resize for emoji use. You still write the animation logic, which is a better fit for custom work than for anyone expecting a gallery of finished assets.
The validator's pass result is based on dimensions. It reports file size, frame count, duration, and estimated frame rate, but is_slack_ready returns the dimension check. That distinction matters when a tool name sounds broader than its decision. The skill itself recommends keeping emoji animation under 3 seconds and using 48 to 128 colors, so a production check should enforce your chosen size and duration limits instead of relying on one Boolean.
Four document skills use different terms from the open examples
The root README says many skills are available under Apache 2.0, then separates the DOCX, PDF, PPTX, and XLSX folders as source available. GitHub supplies no repository-wide license identifier. Copying one of those folders into a company toolkit therefore starts with its local license file, not with an assumption about the organization or the rest of the tree. Anthropic also labels every skill here as demonstration and educational material.
Open issue 1943 adds a current installation warning. The reporter used Claude Desktop 2.16120.0 on Windows 11 with Claude Code 2.1.286 and saw conflicting plugin manifests after marketplace refreshes. Six refreshes reportedly reproduced the problem for three installed plugins at commit 8a1541c; updating from the CLI restored them until the next refresh. This is one reported environment, not proof that every install fails, but affected Windows teams should test refresh behavior before rollout.
A September 29 push shows activity despite 1,392 open items
GitHub recorded 179,344 stars and a last push on September 29, 2026. On October 2, the open queue split into 390 issues and 1,002 pull requests. Recent issue and pull-request updates show that people are still reporting, proposing, and merging work. GitHub returned no latest release, however, so adopters cannot use a tagged release line as their upgrade boundary. Pin a commit and review the diff when moving it.
The practical purchase decision is narrow. A 23-second install and 13-second build make the GIF example easy to inspect, while 0 collected tests leave its behavior for you to prove. Choose one skill, audit every script and permission it can reach, then add fixtures around the outputs your team cares about. If you need a catalog that arrives with repository-wide CI and versioned releases, this one does not meet that bar.