The Developer Pulse #3 — Skills Are the New Apps
Every Monday, Wednesday, and Friday, GitHub trending gets distilled into the one thesis that matters. I read the repos so you don't have to — then surface the pattern the hype cycle missed. Join 50+...
Every Monday, Wednesday, and Friday, GitHub trending gets distilled into the one thesis that matters. I read the repos so you don't have to — then surface the pattern the hype cycle missed. Join 50+ engineers, investors, and tool-builders who read it first.
Today: Two anti-slop repos are trending simultaneously. The "skills" ecosystem for AI agents is adding stars faster than any application framework. The platform war is over — the new battlefield is taste.
Two repos caught my attention this morning: stop-slop (5,181 stars, +539 today) and taste-skill (22,611 stars, +1,430 today). They're not libraries. They're not frameworks. They're personality files — and they're out-starring actual applications.
When I looked at the top 10 trending repos today, four of them are agent skill packs:
- taste-skill (22.6K stars) — "gives your AI good taste, stops it from generating boring, generic slop"
- ECC (194K stars) — "skills, instincts, memory, security" — an agent harness that works across Claude Code, Codex, OpenCode, and Cursor
- Anthropic Cybersecurity Skills (10.3K stars) — 754 structured skills mapped to MITRE ATT&CK, NIST CSF, D3FEND
- stop-slop (5.1K stars) — a single skill file for removing AI tells from prose
This is not a coincidence. The market is voting: the scarce resource is no longer compute or models — it's taste encoded as portable agent configuration.
The ECC Signal
affaan-m/ECC at 194,913 stars is the elephant in the room. It describes itself as "the agent harness performance optimization system" — skills, instincts, memory, and security for any coding agent. Not a new idea. But 194K stars says it's not the idea that matters — it's the format. Developers want their agent configuration to be like their dotfiles: portable, composable, and independent of any single platform.
When ECC advertises compatibility with Claude Code, Codex, OpenCode, Cursor, and "beyond," that's not a feature list. That's the product. The platform is irrelevant. The skills are the asset.
One Number That Tells the Story
| Category | Combined Stars Today |
|---|---|
| Agent skills (taste + ECC + stop-slop + cybersecurity) | ~199,000 |
| Actual applications (Jellyfin + FreeDomain + awesome-free-apps) | ~73,500 |
The skill layer is adding stars at nearly 3x the rate of the application layer. Last week I wrote about infrastructure unbundling. This week the data is sharper: the unbundling is happening at the personality layer, not the API layer.
What Anthropic Knows
Anthropic's own knowledge-work-plugins repo (16,843 stars, +1,718 today) is sitting at #4. That's Anthropic — the model company — publishing open-source agent skills. They understand what the open-source community already proved with ECC and taste-skill: whoever owns the skill format owns the distribution channel.
If your agent's skills are locked to your platform, you own the developer. If skills are portable, the developer owns themselves. Anthropic publishing open skill plugins is either generous open-source stewardship or the smartest platform play in AI right now. Probably both.
Overhyped
- "The best model" debates. The models are within 3% of each other on every benchmark. The skill layer — taste, memory, instincts — has 100x more impact on agent output quality than switching from one frontier model to another.
- Single-agent workflows. ECC has 194K stars precisely because it works everywhere. Building for one agent is like building an iOS-only app in 2016. Possible. Strategic mistake.
- "Prompt engineering." Stop-slop, taste-skill, and cybersecurity skills are curriculum, not prompts. They're persistent, version-controlled, and shared across teams. Prompt engineering was the prototype. Skills are the product.
Prediction #2 — Track This
Within 6 months, a major VC firm will publish an "AI Skills" market map — and the leading category won't be "prompt engineering," it'll be "agent personality infrastructure."
I'll grade this publicly. ECC at 194K stars and taste-skill at 22K are the early data points. The market map is already being drawn — the analysts just haven't named it yet.
Friday: what happens when agent skills become a marketplace — and who controls the transaction.