Summary

A walkthrough of the official Claude Skills leaderboard, which ranks all 85,000+ published skills by install count. The presenter counts down the twelve most-installed, from Lark Doc at 560,000 to Vercel Labs' Find Skills at 2.9 million, explaining what each does and demonstrating three of them live. The framing throughout is that Claude Skills had no discovery mechanism — no equivalent of an App Store top-charts page — until this leaderboard surfaced.

• Find Skills (Vercel Labs, 2.9m installs) is the runaway leader, roughly 3.5x the next entry. It is a meta-skill: you describe what you want in plain language and it searches the catalogue and installs the match. The presenter calls it 'the front door to everything else on the list' and demonstrates it finding a Face Swap skill (342,000 installs) from a spoken request.
• Six of the top twelve are by one author, Matt Pocock — Triage (567,500), Handoff (569,500), a bundle installer for his own skills (606,000), TDD (just under 626,000), Improve Codebase Architecture (682,000) and Grill Me (831,000), plus a Grill With Docs variant. A single independent author holds half the leaderboard.
• Handoff addresses what the presenter calls 'context rot' — long sessions becoming forgetful. It writes a compact Markdown file carrying the state, decisions and next steps out of an exhausted session so a fresh one can resume. He demonstrates it on a real project of his own and likens it to a hospital shift change.
• Grill Me (831,000) makes Claude interrogate a plan before building it. Demonstrated on an AI-thumbnail SaaS idea, it returns eight questions covering who pays, what the landing page promises, the wedge into a crowded market, proof of demand, and an explicit kill switch.
• TDD reframes test-driven development as a general AI-work principle rather than a coding one: define what 'correct' looks like before Claude produces anything. The presenter ties this directly to AI output 'that looks right but quietly isn't' — his stated reason for AI slop.
• Two Anthropic and Vercel official skills sit high: Anthropic's Front End Design (767,200; 168,000 GitHub stars) exists to stop generated interfaces converging on the same fonts, purple gradients and layouts, and Vercel's Agent Browser (40,000+ stars) gives an agent a real browser to open pages, click, and fill forms.
• The Agent Browser demonstration is an autonomous purchase: the presenter instructs it to 'buy everything I've ever bought from Amazon again' and it opens Chrome and begins adding past orders to the cart, unattended.
• Caveat on the ordering: the video's own countdown does not stay consistent. Grill Me at 831,000 is revealed while discussing the number-four slot, and the number-three entry (767,200) is reached afterwards. The install figures are reliable; the rank positions between 2 and 4 are not.

Why it matters

Two things worth carrying forward. First, discovery is now the highest-value capability in the skills ecosystem by a wide margin — the tool for finding tools out-installs every tool that does a job by roughly 3.5x, which says the catalogue has outgrown any manual way of navigating it. Second, the most-installed skills are overwhelmingly about process discipline rather than new capability: grill the plan before building, define correct before producing, hand context over cleanly, decide what to do first. That is the same pattern showing up across independent sources — the constraint on useful AI work is the discipline around it, not the model.

SX7WbJnQeSM-transcript.txt