Open GitHub Trending today and 4 of the top 10 are "skill packages."
obra/superpowers, 199,491 stars. multica-ai/andrej-karpathy-skills, 139,768 stars. Imbad0202/academic-research-skills, 15,192 stars. Add HKUDS/CLI-Anything — a "make ALL software Agent-Native" framework — at 38,237 stars.
In a single day, superpowers gained 1,776 stars, karpathy-skills gained 2,620. The academic research workflow repo gained 1,639.
The numbers are pretty. But what do they actually represent?
What Are "Skill Packages"?
Simply put, they're CLAUDE.md files (or sets of files) filled with prompts and behavioral instructions for Claude Code. They tell the AI how to modify code, how to do reviews, how to write academic papers, how to make architectural decisions.
obra's superpowers calls itself "an agentic skills framework & software development methodology that works." karpathy-skills' README is one line: "A single CLAUDE.md file to improve Claude Code behavior, derived from Andrej Karpathy's observations on LLM coding pitfalls."
At their core, these are systematized prompt engineering. Taking AI coding experience scattered across blogs, tweets, and forums, and organizing it into structured configuration files that Claude Code follows.
This has value. Good prompts do significantly improve AI coding assistant output quality. But 199K stars?
Star Inflation
Let me be direct: a single markdown file with 199K stars — that number itself deserves skepticism.
Not that it has no value. But stars in this context have lost their original signal meaning.
The traditional meaning of a GitHub star is: "this project is worth watching" — implying it's runnable code, a useful tool, an implementation worth studying. But a CLAUDE.md file isn't code. It's a configuration file. You fork it into your project and it just sits there. You don't "use" it; you just "own" it.
Starring these repos feels more like making a statement — "I believe AI-assisted programming is the future" — than "this tool is useful to me."
That's the problem. When stars become position statements, they stop being quality signals.
Another Angle: This Is Actually a Good Thing
That said.
199K people starring a prompt configuration file — regardless of how you interpret the behavior itself — demonstrates one thing: demand for AI-assisted programming is spilling over from the developer community into much broader audiences.
Before, only hardcore programmers starred things on GitHub. Now, a data analyst using Claude Code to write Python scripts, a graduate student using AI to write a thesis, a marketer using vibe coding to build landing pages — they're all starring these repos.
This is an expansion of GitHub's community boundaries. It dilutes the gold content of stars, but expands the audience for AI programming tools.
From an industry development perspective, the latter matters much more than the former.
CLI-Anything and the "Agent-Native" Narrative
HKUDS's CLI-Anything deserves separate mention. Its ambition is bigger — not making a skill, but building a Hub for "making ALL software Agent-Native."
The project, initiated by Hong Kong University's Data Science Laboratory, has already reached 38K stars. The website is clianything.cc.
"Agent-Native" is becoming a new technology narrative. Like "Cloud-Native" in the 2020s — it doesn't describe a specific technology, but a directional judgment: future software should be designed for agents, not for humans.
This judgment is probably correct. But between "probably correct" and "usable right now" lies a massive engineering gap.
My Take
The star surge in the skills ecosystem is driven by three factors:
- Claude Code user growth — more users means more shares and stars naturally
- Productization of prompt engineering — organizing scattered experience into distributable packages is a real product form
- Stars as community identity — in the early stages of AI programming, starring a popular skill repo is saying "I'm part of this community"
I use Claude Code myself. I do use some community skills. They're useful. But that 199K star number — I'd suggest looking at it with a 50% discount.
What really matters isn't the star count. It's: what are the actual adoption and retention rates of these skills? How many people starred and actually used them for more than a week? How many use both superpowers AND karpathy-skills simultaneously? How do conflicts between these skill sets get resolved?
Those questions are the real health indicators of this ecosystem. Star counts are marketing metrics, not health metrics.
Main sources: