Anthropic used to publish skills as readable, forkable building blocks. Now some of its most powerful ones - Ultraplan, Ultrareview, Cloud Security - ship only inside Claude's paid cloud tier, where you can use them but can't read what they actually do.
That distinction matters more than it might sound.
A skill, in Claude's architecture, is a prompt bundle that tells the model how to approach a specific type of task: code review, planning, security auditing. The value wasn't just in having them. It was that you could read them, combine them, and wire them into your own pipelines. A developer could take a review skill, pair it with a language-specific agent, wrap both in a workflow that saves outputs at each stage, and debug exactly what went wrong when it failed. The prompt was the interface.
With cloud-gated skills, that's gone. You get the capability but not the machinery. You can run Ultrareview, but you can't see what it's checking for, can't modify it for your stack, and can't trace the logic when it gives you a wrong answer.
What Composability Actually Means Here
A recent analysis at vexjoy.com puts it plainly: "You can't compose what you can't read, and you can't diagnose what you can't inspect."
This isn't abstract. Take a practical code review pipeline: one skill reads the diff, another checks for security patterns, a third writes the PR comment, and each phase saves an artifact so you can see where the chain broke. That's only possible because you could read each piece. When the security check missed something, you opened the skill, saw it wasn't testing for SQL injection in parameterized queries, and fixed it. With opaque cloud skills, your debugging path is "run it again and hope."
Composability - the ability to connect small, readable components into larger auditable systems - is how serious Claude users have been building production workflows. Gating the newest skills behind a paywall doesn't just restrict access. It changes the fundamental contract: these are now third-party services, not building blocks.
The Incentive Behind the Shift
Anthropic's reasoning isn't hard to read. Ultraplan and Ultrareview are differentiated products, and cloud gating lets the company monetize them directly while iterating faster without worrying about breaking thousands of forked community versions. That's a defensible business decision.
But the cost is real for developers who adopted Claude Code specifically because they could inspect and modify skills. The most capable new additions are no longer available on those terms.
The pattern is familiar from other developer platforms: open primitives drive adoption, then premium features get locked into subscriptions. The transition isn't inherently bad, but it means the community needs to be clear-eyed about which capabilities are actually composable and which are vendor-managed black boxes with a skills-shaped interface.
For practitioners who have built pipelines around open skills: audit what you're relying on now. If a skill you depend on is still open, consider pinning or forking it. For the cloud-gated skills, treat them like any third-party API - useful for the task, but not a foundation you can build transparent, auditable workflows on top of.