What Are Agent Skills? Inside GitHub's Fastest-Growing Trend (June 2026)

TL;DR: "Agent skills" — modular folders of instructions, scripts, and references that teach AI coding agents how to do specific jobs — are the fastest-growing repo category on GitHub right now. GitTube's radar flagged five skills repos in 72 hours, including one accelerating at 1,035 stars-velocity units the morning this was published and another at 2,526 — the highest acceleration in the entire 3,300-repo index. Here's what they are, why the wave is happening now, and which repos are riding it.

What exactly is an agent skill?

An agent skill is a packaged unit of procedural knowledge for an AI agent — typically a folder with a SKILL.md file describing when and how to perform a task, plus optional scripts and reference files the agent loads on demand. The format was popularized by Anthropic's Claude Code, and the idea generalizes across agents: instead of pasting the same instructions into every prompt, you install a skill once and the agent discovers and applies it when relevant.

Think of it as the npm moment for agent behavior. Code got package managers in the 2010s; agent know-how is getting them now. A skill for "review my PRs like a staff engineer" or "write product requirements docs" is suddenly something you git clone rather than re-prompt.

That's why this matters beyond the hype: skills are composable, shareable, and forkable — exactly the properties that made open source compound. Whoever owns the popular skill registries owns real distribution to developers.

The wave, in numbers (live radar data)

These aren't impressions — they're snapshots from GitTube's live leaderboard, which tracks 3,300+ repos every 30 minutes and scores acceleration before repos hit the newsletters:

Live proof — these badges render straight from the radar and update every 30 minutes:

And this isn't a 72-hour blip: the radar has been tracking the skills cluster since spring — andrej-karpathy-skills hit 16K stars in its viral ignition window in early May, and awesome-codex-skills fired extreme breakout signals back in late April. When five more repos in one category fire within 72 hours — including a Google org repo — that's not a coincidence. That's a category forming. The same multi-repo pattern preceded the AI-coding-assistant wave and the MCP-server wave (which went from novelty to dozens of clones of every popular server in months).

Why now? Three forces stacking

1. Agents stopped being chatbots. Claude Code, Cursor agents, and autonomous pipelines now run long, multi-step jobs. Long jobs need procedure, not vibes — and procedure wants to be versioned, reviewed, and shared. A markdown file in a repo beats a prompt in someone's head.

2. The skill format is trivially forkable. A skill is mostly prose plus a little glue code. The barrier to contributing is the lowest of any "AI infrastructure" category ever — which is rocket fuel for star velocity. Compare: writing an MCP server takes a weekend; writing a useful skill takes an hour.

3. Non-engineers just got a door in. Look at what's actually breaking out: pm-skills — product-manager skills. The wave's freshest repo isn't for coders, it's for the people working next to coders. When a developer-tool category crosses into adjacent professions, its ceiling multiplies — the crossover pattern trend-spotters like Exploding Topics built a business on, and the bottom-up adoption motion a16z has long argued is the strongest developer GTM there is.

How to ride a wave like this (before the newsletters)

By the time TLDR or Changelog Nightly covers a repo, the first-mover window for content, integrations, and positioning is mostly gone. The compounding move is catching the acceleration, not the star count:

1. Watch acceleration, not stars. A repo at 4,000 stars gaining 1,200/day beats a 100K-star monument. GitTube's leaderboard sorts by predictive viral potential for exactly this reason — and GitHub's own Trending page has reliability problems plus a well-documented fake-star economy, so raw star counts mislead.
2. Look for the cluster. One rising repo is a project; three in the same category is a market. The skills cluster qualified months ago and re-confirmed this week.
3. Ship something into the wave while it's cresting — a skill, a comparison, an integration. Speed beats polish here; the conversation is already at peak attention.

Maintainer of one of these repos? Claim your listing and embed your live badge — it updates itself every 30 minutes. And if you want the radar to catch the next wave hours earlier, lend it a read-only GitHub token — every token adds 5,000 requests/hour of scanning capacity.

The meta-lesson

Every page on this site, including this one, is structured data plus fresh proprietary numbers — which is precisely what AI search engines cite. Perplexity indexes new tech content within hours; ChatGPT within days. If you ship a product and want it visible where developers actually search now — inside AI assistants — check what those engines see of you: run a free AI-visibility scan on LoudPixel. It scores your site the way we score repos: with numbers, before your competitors get there.

Data: GitTube radar, 2026-06-10. Acceleration = change in star velocity between snapshots; z-score = how unusual current velocity is vs the repo's own history. Updated every 30 minutes from 43,000+ snapshots.