What's this about?
Anthropic open-sourced a `/claude-api` skill and built it into Claude Code. It is not a standalone tool but a set of structured reference documents and workflow instructions: when Claude Code detects an Anthropic SDK import in your project — or you type `/claude-api` — it loads the API documentation for your language into context. Eight languages (Python, TypeScript, Java, Go, Ruby, C#, PHP, cURL) across two surfaces: the Messages API (requests, streaming, tool calls, batching, prompt caching) and Managed Agents (Anthropic-hosted stateful agents with sandboxed environments).
Why ship documentation as a skill rather than training on it?
Because APIs change faster than models are trained. A model's knowledge of an SDK is frozen at its training cutoff, which is why it confidently writes calls against last year's signatures. Loading the current documentation into context at the moment it is needed sidesteps that entirely — and it is a much better answer than hoping the next training run includes the changelog.
Why does this matter to you?
- This is the general answer to model staleness for any API, and it is a pattern worth copying for your own.
- It only triggers on SDK detection or explicit invocation, so it costs nothing when irrelevant.
- dotey's caveat: if you are only a Claude Code user rather than building against the API, this is not for you.