The API

Implementation guides for building directly against the Claude API: tool use, streaming, structured outputs, error handling, prompt caching, batching, and the SDK-level details that don't show up until you're running something in production. Each guide names the specific request or response shape involved rather than describing the feature abstractly, and links out to the relevant entry in the error catalogue wherever a mistake here produces a specific, greppable error.

Most of what breaks in a real integration isn't the basic "send a message, get a response" path — that part is well documented and rarely surprises anyone. It's the second layer: what happens when a stream disconnects mid-response, how a tool-use loop should handle a tool that fails, what a batch job's completion actually looks like versus what you assumed it would look like, and which parameters quietly stopped being accepted between one model generation and the next. This cluster is built around that second layer specifically, because it's where production code actually spends its debugging time.

If you're migrating between model generations rather than building something new, the prompting cluster covers what changed in request shape and default behaviour more directly than these implementation guides do — the two clusters are complementary, not overlapping.