Anthropic just formalized what power users have been hacking together in Markdown files for months—and the open-source response is already bigger than the official product.
The Summary
- Anthropic released Claude Code Routines, official structured instruction packages for coding agents, while an open-source project called claude-skills hit 5,200+ GitHub stars with 232+ production-ready skills that work across 12 different AI coding tools
- The gap between "official feature" and "what developers actually built" reveals how fast the agent tooling layer is moving—faster than any single company can standardize
- Both approaches solve the same problem: AI coding agents are general-purpose, but real work requires domain expertise in DevOps, compliance, marketing, or C-suite advisory
The Signal
Anthropic's Claude Code Routines are structured instruction sets that give coding agents repeatable workflows. Instead of re-explaining your deployment process or compliance requirements every session, you package the knowledge once. The Hacker News discussion pulled 254 points and 156 comments, which means this hit a nerve with people who've been rebuilding the same context windows for months.
But the real story is what happened in parallel. The claude-skills repository is already shipping 232+ skills with 305 Python CLI tools—all standard library, zero pip dependencies. Every skill includes a SKILL.md file with instructions, Python tooling, and reference docs. It's not just for Claude Code. The same skills run on OpenAI Codex, Gemini CLI, Cursor, Aider, Windsurf, and seven other platforms via a conversion script.
"One repo, eleven platforms. This is what interoperability looks like when companies won't give it to you."
Three details matter here:
- Scope beyond code: Skills cover engineering, DevOps, marketing, compliance, and C-level advisory. This isn't just "write a function." It's "help me navigate SOC 2 attestation" or "draft a product roadmap."
- Zero external dependencies: All 305 Python tools use only the standard library. That means they run anywhere, no environment setup, no version conflicts.
- Portability by design: The conversion script turns Claude-native skills into formats for Codex, Gemini, Cursor, and others. The community is standardizing faster than the vendors.
Anthropic is doing what big AI labs do: release a polished feature for their ecosystem. But developers are doing what developers do: build the tool they actually need, make it work everywhere, and open-source it before the official version even has a pricing page. The GitHub repo already has more stars than most venture-backed developer tools get in a year.
This isn't just about Claude. It's about the agent skills layer becoming the new middleware. In Web2, we had API wrappers and SDKs. In Web4, we have skill libraries and instruction packages. The value isn't in the base model anymore—it's in the domain-specific context you can reliably inject.
The Implication
If you're building with AI coding agents, stop waiting for vendors to ship the perfect plugin system. The open-source skills layer is already here, and it's platform-agnostic by necessity. Download the claude-skills repo, fork what you need, and contribute what you build. The companies will eventually adopt whatever becomes standard in the wild.
For companies: your competitive advantage isn't the model you use. It's the skills library you maintain. Start treating agent instructions like infrastructure. Version them, test them, and document them like you would any other critical system.