Build apps with the Claude API or Anthropic SDK. TRIGGER when: code imports `anthropic`/`@anthropic-ai/sdk`/`claude_agent_sdk`, or user asks to use Claude API, Anthropic SDKs, or Agent SDK. DO NOT TRIGGER when: code imports `openai`/other AI SDK, general programming, or ML/data-science tasks.

Building LLM-Powered Applications with Claude

Choose the right surface, detect the project language, read the relevant docs.

What You Get

Language-detected SDK examples (Python, TypeScript, Go, Elixir, Rust, Lua, cURL)
Decision tree: single call vs workflow vs agent vs Agent SDK
Current model IDs, thinking/effort config, caching patterns
Tool use patterns (tool runner, manual loop, code execution)
Error taxonomy and live documentation URLs

Top Pitfalls (quick reference)

Mistake	Fix
Using `budget_tokens` on Opus 4.6 / Sonnet 4.6	Use `thinking: {type: "adaptive"}` -- budget_tokens is deprecated
Lowballing `max_tokens`	Default ~16K non-streaming, ~64K streaming. Truncation = wasted retry
Prefilling assistant message on Opus 4.6	Returns 400. Use structured outputs or system prompt instead
Silent cache invalidation	Check `usage.cache_read_input_tokens`. Common culprit: `datetime.now()` in system prompt
Redefining SDK types	Use `Anthropic.MessageParam`, `Anthropic.Tool`, etc. -- don't roll your own

Full list: shared/pitfalls.md

Defaults

Use Claude Opus 4.6 (claude-opus-4-6) unless the user names a different model. Use adaptive thinking (thinking: {type: "adaptive"}) for anything remotely complicated. Use streaming for long input/output -- .get_final_message() / .finalMessage() if you don't need individual events.

Language Detection

Infer from project files:

*.py, pyproject.toml, requirements.txt -> Python -- read from python/
*.ts, *.tsx, package.json, tsconfig.json -> TypeScript -- read from typescript/
*.js, *.jsx (no .ts present) -> TypeScript -- JS uses the same SDK
*.go, go.mod -> Go -- read from go/
*.ex, *.exs, mix.exs -> Elixir -- read from elixir/
*.rs, Cargo.toml -> Rust -- read from rust/
*.lua, .luarc.json -> Lua -- read from lua/

If ambiguous, ask. If unsupported language, suggest curl/ examples.

Which Surface?

Use Case	Surface
Single call (classify, summarize)	Claude API
Multi-step pipeline, your tools	Claude API + tool use
Agent needing file/web/terminal	Agent SDK
Custom agent, your own tools	Claude API agentic loop

Decision tree and "Should I Build an Agent?" criteria: shared/surfaces.md

Reading Guide

Task	Read
Single call (classify/summarize/extract)	`{lang}/claude-api/README.md`
Chat UI / streaming	+ `{lang}/claude-api/streaming.md`
Long conversations (context overflow)	+ Compaction section; details: `shared/thinking-effort.md`
Prompt caching / cache optimization	`shared/prompt-caching.md` + `{lang}/claude-api/README.md`
Tool use / function calling / agents	+ `shared/tool-use-concepts.md` + `{lang}/claude-api/tool-use.md`
Batch processing	+ `{lang}/claude-api/batches.md`
File uploads across requests	+ `{lang}/claude-api/files-api.md`
Agent with built-in tools	`{lang}/agent-sdk/README.md` + `{lang}/agent-sdk/patterns.md`
Model selection / capabilities	`shared/models.md`
Thinking, effort, compaction	`shared/thinking-effort.md`
Error handling	`shared/error-codes.md`

Go, Elixir, Rust, Lua, cURL: single file covers all basics. Read that plus shared/ files as needed.

When to Use WebFetch

Use when user asks for "latest" info, cached data seems wrong, or features aren't covered here. URLs: shared/live-sources.md.

claude-api