Tell us what you want to make. We'll show the evidence-backed path.

Current evidence-backed paths

Three starting points, each labeled by what the existing runs really prove. Inspect the final evidence and download a reproduction pack before installing anything.

Every skill we've tested

Sorted by stars; every number generated from the raw test logs, last test dated in each record. Click through for the full write-up — environment, timings, and the failure log.

Why this exists

Claude Code skills exploded in 2026, and so did the listings of them. What none of the lists tell you: does npx skills add actually work for this repo? How long does a 51.1MB skill take to land? Does it run on Linux, or does it quietly need a Mac with a desktop app open? Which famous repo is actually archived while every directory still links to it? (One of the biggest ones, it turns out.)

Those answers only come from running the install. So that is the whole editorial policy here: no page ships until we have run the thing. The test record at the top of each page — repo, commit, environment, timings, failures — is generated straight from structured test records. Narrative and historical observations are dated separately; summary metrics are never copied into prose by hand.

How to install Claude Code skills (the short version)

The ecosystem has largely converged on one installer, the skills CLI. For nearly every package we have tested, this is the whole story:

npx skills add <owner>/<repo> -a claude-code -y

First run pays a few seconds of CLI bootstrap; after that installs are typically two to three seconds. Skills land in .claude/skills/ in your project (add -g for user-level). The interesting failures happen after install — at use time, when a skill turns out to need a browser binary, a desktop app, or a rendering pipeline. That is exactly what each test page documents.

Beyond install tests: deep comparisons

Install tests tell you a skill runs. Deep comparisons tell you which one to use: we give two competing skills the identical brief, follow each one's own methodology, render both results side by side, and file improvement suggestions with both authors. Two so far: video-shotcraft vs Remotion's official skills — same five-second scene, two very different theories of what an agent needs — and the three design skills pointed at one real page: ours, with the findings against ourselves published.

We also publish intent-specific decision pages only when an existing test can support them. The first pilot is the tested alternative to Remotion's official skills; it reuses the same measured record instead of inventing a new comparison.

Guides from the same logs

The test sessions also produced four guides: eight best practices we learned the empirical way, how to make skills actually activate, how to build a skill (anatomy of 8 real ones), and a short answer to "what are Claude skills". Every claim in them links back to a test record.

Quick answers

What's next

More community skills enter the test queue weekly, and everything already tested goes on a 30-day retest calendar — repos move, break, and get archived (ask GSD). Watch the changelog.