The GitHub trending list is dominated by "skills" packs that vary 1000x in scope but all use Anthropic's SKILL.md format, which is becoming a real distribution primitive for AI agent capabilities. The article analyzes the implications for builders, including portability challenges, the split between workflow enforcement and capability extension, and the low moat for skills packs.
Three of the trending "skills" repos, read for what's inside: * karpathy-skills: 1 markdown file, 4 rules, \~70k stars * mattpocock/skills: \~10 SKILL.md files, \~115k stars * everything-claude-code: 182 SKILL.md files + 48 agents + 68 commands + hooks + rules + MCP configs + npm packages, \~178k stars (plus a second repo by the same author also in the top 10) Three orders of magnitude in scope, same label. What's underneath all of them is Anthropic's SKILL.md format. Markdown with YAML frontmatter, auto-loaded by Claude Code (and via shims by Cursor, Codex, OpenCode, Gemini, Antigravity) at session start. It's prompt-with-conventions. That's the actually interesting part. For two years, the answer to "how do I make this agent better at X" was prompt engineering, manual context, glue code. SKILL .md is the first widely-adopted attempt to make a unit of agent capability publishable, forkable, and installable, with a defined invocation pattern (file gets auto-loaded, frontmatter declares when it fires). That's a real primitive even if the trending list around it is noisy. Four implications I keep coming back to as a builder: 1. The packaging is the primitive, not the content. mattpocock's tdd skill and ECC's tdd-workflow skill solve the same problem with similar prose. The differentiation is whether you ship it in a format other people can compose with, not whether your wording is cleverer. 2. Cross-harness portability is leaky. ECC ships into .claude/, .cursor/, .codex/, .opencode/, .gemini/, .agent/, six paths, each with its own quirks (Cursor has 20 hook events vs Claude Code's 8, plugin distribution can't carry rules, OpenCode has a different plugin system). Write-once-run-anywhere is real-ish but you pay for it. 3. Workflow enforcement vs capability extension is a real split. Most of the agent-skills discussion in 2024-2025 was capability (tool use, browser, APIs). What's actually trending in mid-2026 is mostly workflow (TDD, triage, code review, planning). Different bet about where value lives. 4. The bear case. If your "skill" is a markdown file derived from a tweet and it has 70k stars, the moat is roughly zero. A startup whose differentiation is a skills pack should assume that pack gets copied, forked, or absorbed into the base model within 12 months. Karpathy's four rules will be in the default behavior of the next Claude release. For people shipping production agents, are you treating skills as a real distribution primitive (publishing, versioning, dep-managing), or as personal scratch that occasionally gets pushed to GitHub? And does the cross-harness story hold up for you or do you end up forking per setup?
A tweet thread curating 11 open-source GitHub repositories for AI tools, agents, and learning resources, including PilotDeck, Karpathy's skills, and Microsoft's AI agent course.
SkillKit is a package manager for AI coding agent skills, supporting 46 agents and 400K+ skills from 31 sources, allowing one skill to be used across multiple agents.
A LinkedIn post shares 10 GitHub repositories covering AI agent skills, LLMs, prompt engineering, and generative AI, including free courses and practical resources.
Curated GitHub repository offering 1100+ real-world AI agent skills from major dev teams like Anthropic, Google, Stripe, and Vercel, compatible with Claude Code, Codex, Cursor, and other AI coding assistants.
The article reviews the top five AI Agent projects on GitHub with the fastest star growth recently, highlighting Anthropic's Financial Services Agent Framework, ByteDance's UI-TARS Desktop, and various coding Agent tools.