Claude API Builder — the three things most Anthropic SDK code skips

A working Claude integration and a good one are separated by three habits, and almost every first draft skips all three.

Prompt caching. A 4,000-token system prompt re-sent uncached on every request is the most common — and most expensive — mistake in production LLM code. Mark the stable prefix with a cache breakpoint and the cost and latency drop immediately. It's the single biggest win and the most-skipped step.

Streaming. Blocking calls with high max_tokens time out under load. Streaming prevents it, and the SDK's final-message helper gives you the whole response back if you don't need to handle individual events.

Current model defaults. Code that hardcodes a model version that's since retired breaks quietly. Pin to the current Opus model, default to adaptive thinking for anything non-trivial.

What the skill does

Takes a feature you want to build — chat, tool use, structured output, batching — and produces working Anthropic SDK code in your project's language with all three habits built in. It detects the language, refuses to mix providers (it won't write Anthropic calls into a file that imports OpenAI), and won't guess SDK method names it can't verify. It also migrates existing Claude code between model versions.

Part of a 26-skill drop

This is one of twenty-six new skills that landed today, bringing /skills to 200. The developer pack got the most: alongside this one, a Browser Automation Driver, React performance and composition reviewers, a Postgres performance reviewer, test-driven-development, and a codebase architecture reviewer. Marketers, web builders, and founders each got their own additions.

Every one is a self-contained SKILL.md you can install and read in a minute — and each was checked against the existing library so none of them duplicate what was already there.

Full skill — with the provider-check, caching breakpoint guidance, and the streaming and model defaults spelled out.