Agent skills are still new enough that no single approach has won. The providers building on the format are making different bets about what matters most — domain coverage, distribution, interoperability, operational depth. Surveying their repositories is one of the fastest ways to understand where this is all going.
I went through the public skills repos for Anthropic, Vercel, Microsoft, Speakeasy, OpenAI, Hugging Face, Cloudflare, and Postman this week. Here’s what I found.
Anthropic: Category Depth and the Meta-Skill
Anthropic’s skills repo is organized by category and then by language — a nested structure that signals they’re thinking about this as a real taxonomy problem, not just a collection of scripts. They’ve got categories ranging from algorithmic art to cloud API to document generation.
The standout is their skill creator skill — a meta-skill for generating new skills. I always pay attention to the recursive ones. If the tool can build itself, that tells you something about the maturity of the abstraction. It also points toward a future where skill generation is automated, which has obvious implications for how organizations scale their skill libraries.
They’ve also connected their repo to the Claude plugin marketplace, which is the distribution layer I’m most interested in. A skill that lives only in a local repo is a draft. A skill in a marketplace is infrastructure.
Vercel: Domain Focus Done Simply
Vercel’s approach is straightforward: a small set of skills tightly scoped to their domain. React best practices, web design guidelines, React Native, composition patterns. No sprawl.
This is a reasonable starting point for a platform with a well-defined developer audience. You know who you’re serving and what they need. The risk is that focused skill sets calcify — as the platform grows, the taxonomy needs to grow with it. But as a model for “start with what your users actually need,” Vercel’s repo is worth studying.
Microsoft: The Plugin Model and the VS Code Integration
Microsoft is doing something architecturally different. Their skills repo isn’t just a collection of markdown folders — it’s organized around a plugin model where curated sets of skills are bundled and installed via the Copilot CLI.
This matters for governance. Instead of individual developers pulling whatever skills they want, the plugin model lets organizations distribute sanctioned skill bundles. It’s closer to a package manager than a file share.
They’ve also integrated MCP server configurations directly into VS Code settings. Their categorization — docs, development, and utilities — maps cleanly onto how developers actually spend their time. Docs MCP servers include Microsoft Learn and Deep Wiki. Development includes GitHub, Playwright, and Terraform. Utilities handle sequential thinking and memory.
What I hadn’t seen before was their Sensei-style scoring system for code generation: generate, evaluate, analyze, regenerate, report. It’s a feedback loop baked into the workflow. That’s a more sophisticated approach to skill execution than most repos are showing right now.
Speakeasy: The API Lifecycle as a Skills Framework
Speakeasy is the one I keep coming back to as a model. Their skills are organized around the API lifecycle — SDK configuration, hook customization, code generation — and they’ve built an evaluation harness that creates an isolated workspace, copies the relevant skills into it, runs the agent, and scores the output.
That evaluation layer is what elevates their repo above the others. It’s not enough to write a skill and ship it. You need to know whether it works. Speakeasy has operationalized that question.
Their repo also makes a strong case for something I’ve been arguing for a while: every stage of the API lifecycle should have a corresponding skill set, published and maintained by the service provider. If you’re an SDK generator, your skills are your developer interface. Speakeasy gets this.
OpenAI (Codex): Third-Party Curation and the Docs-as-Skill Pattern
Codex distinguishes between curated skills (third-party integrations) and system skills (core infrastructure). Their curated list includes Cloudflare, Figma, GitHub, Jupyter, Netlify, and Notion — which starts to look like a directory of the modern developer stack.
The most interesting item in their repo is their own docs surfaced as a skill. This closes the loop on what I’ve been exploring with MCP: documentation isn’t just a website you point developers to. It’s a knowledge artifact that can be packaged, versioned, and loaded by an agent on demand. OpenAI treating their own docs as a first-class skill is a signal worth paying attention to.
Hugging Face: ML Tasks as the Organizing Principle
Hugging Face’s skills are organized around ML tasks — dataset creation, model training, evaluation, publishing. The taxonomy reflects their community’s mental model of the work.
What’s interesting here is the cross-client compatibility story. Their skills are designed to work across Codex, Gemini, and Cursor — not just one agent environment. That interoperability focus is ahead of where most providers are right now, and it points toward a future where skills travel with the developer rather than being locked to a single platform.
Cloudflare: Operational Depth Over Documentation Breadth
Cloudflare’s skills repo reflects their MCP work in a useful way. They’ve got skills for Workers, durable objects, the agents SDK, Wrangler, and web performance — all operational, all tied to how you actually build on their platform.
They also surface a subset of their MCP servers alongside the skills: docs, bindings, builds, and observability. The docs/operational split I noticed in their MCP work shows up here too. Documentation tells you what something does. Operational skills tell you how to run it in production. Cloudflare is one of the few providers building both.
Postman: Full API Workflow in Two Skills
Postman’s repo is lean — two skills covering general Postman usage and API readiness auditing. But the capabilities packed into those two skills are substantial: syncing collections, generating client code, running tests, creating mock servers, publishing documentation, security auditing.
The developer relations team attribution in their README is a small thing that matters. Skills built by people who talk to developers every day tend to reflect what developers actually need, not what product teams imagine they need. That provenance is worth noting.
What the Differences Tell Us
Laying all of these repos side by side reveals a few things:
Distribution is the unsolved problem. Most repos are just GitHub folders. Anthropic has a marketplace, Microsoft has a plugin model, Hugging Face has cross-client compatibility. But there’s no standard answer yet for how skills get from a repository to an active developer’s workspace and stay current when they’re updated.
Domain specificity is underutilized. Microsoft’s domain-specific skills language and Cloudflare’s operational depth are the exceptions. Most repos are still general-purpose. The real value of skills — domain expertise that an agent can load on demand — is mostly unrealized.
Evaluation is rare but important. Speakeasy is nearly alone in having a formal evaluation harness. If skills are going to be trusted with real work, we need standard ways to test them. This is the governance gap hiding in plain sight.
The Venn diagram with MCP and OpenAPI is still being drawn. Skills, MCP servers, and OpenAPI specs are all touching the same territory from different angles. The providers who figure out how these three fit together — and build tooling around that integration — will have a significant advantage.
I’ll be talking to people behind several of these repos on the podcast in the coming weeks. The spec is one thing. The reasoning behind the choices is what I actually want to understand.
This is part of an ongoing series on API governance, developer documentation, and the evolving infrastructure of AI-assisted development. Previous entries covered MCP server distribution strategies from Google, Microsoft, and Cloudflare.
