Claude API

The Claude API skill is the official Anthropic Agent Skill that teaches Claude Code (and other agents) to build, debug, and optimize applications on the Claude API and Anthropic SDK — with prompt caching, adaptive thinking, tool use, and Managed Agents wired in by default.

What is the Claude API skill?

It is a skill from Anthropic's anthropics/skills repository that activates whenever your code imports anthropic or @anthropic-ai/sdk, when you ask for the Claude API, the Anthropic SDK, or Managed Agents, or when you tune a Claude feature such as caching, thinking, tool use, batch, files, citations, or memory. It detects the project language — Python, TypeScript/JavaScript, Java/Kotlin/Scala, Go, Ruby, C#, PHP, or raw cURL — and reads the matching language-specific reference files before writing a single line of code, so it never guesses SDK signatures.

What it does

• Writes Claude API calls through the official Anthropic SDK for your language (or raw HTTP only when you explicitly ask), never an OpenAI-compatible shim.
• Adds prompt caching with correct cache_control breakpoint placement and verifies cache hits via usage.cache_read_input_tokens.
• Defaults to the current model ID claude-opus-4-8 and adaptive thinking (thinking: {type: "adaptive"}), with streaming for long requests.
• Migrates existing code between Claude model versions (4.5 to 4.6, 4.6 to 4.7, 4.7 to 4.8) and replaces retired models, translating deprecated budget_tokens and prefill patterns.
• Implements tool use (user-defined and server-side), structured outputs via output_config.format, batches at 50% cost, the Files API, citations, and server-side compaction.
• Sets up Managed Agents — persisted, versioned agent configs plus per-session containers — using the create-once, reference-by-ID flow.

How it works

Everything routes through the single POST /v1/messages endpoint; tools and output constraints are features of that endpoint, not separate APIs. The skill first scans for non-Anthropic provider markers (import openai, gpt-4, *-openai.py) and refuses to inject Anthropic SDK calls into provider-neutral files. It then reads {lang}/claude-api/README.md first, pulling in shared/prompt-caching.md, shared/tool-use-concepts.md, shared/agent-design.md, shared/model-migration.md, and the Managed Agents references only as the task requires. Live SDK documentation is fetched from shared/live-sources.md whenever a binding is not explicitly documented.

Who it is for

Engineers building LLM-powered applications on Claude — backend developers, AI application teams, and anyone integrating the Anthropic SDK into a Python, TypeScript, Java, Go, Ruby, C#, or PHP codebase, or calling the API over cURL. It is aimed at developers who want correct, current, cache-optimized Claude code rather than hallucinated SDK calls.

What you can build

Classification, summarization, and extraction services; streaming chat UIs; code-orchestrated workflows with tool use; custom agents that run their own loop; batch pipelines; document and report generators using the code-execution sandbox; and server-managed stateful agents with file mounts and an SSE event stream via Managed Agents.

Why it matters

LLMs frequently hallucinate SDK method names, append wrong date suffixes to model IDs, and forget prompt caching — quietly inflating cost and breaking builds. This skill pins the agent to documented APIs, the exact claude-opus-4-8 model string, adaptive thinking, and verified caching, so the Claude code your agent ships actually compiles, caches, and stays current across model migrations.

1. Install via the skills CLI

npx skills add anthropics/skills --skill claude-api

2. Or add it as a Claude Code plugin

Register the Anthropic marketplace, then install the example skills plugin:

/plugin marketplace add anthropics/skills
/plugin install example-skills@anthropic-agent-skills

3. Use it

Open a project that imports anthropic or @anthropic-ai/sdk and ask your agent to add a feature — for example, "add prompt caching to this Claude call" or "migrate this file to Opus 4.8". The skill triggers automatically.