AI/ML6 min read

Claude API

Build with the Claude API - messages, system prompts, tool use, streaming, prompt caching, and best practices for production AI applications.

Claude APIAnthropicLLMTool UseStreaming

skill: Claude API category: AI/ML tags: [Claude API, Anthropic, LLM, Tool Use, Streaming] readingTime: 6

Claude is designed to be helpful, harmless, and honest - it works best with clear system prompts, well-structured tool definitions, and explicit output format instructions. Don't fight the model's default behaviors; work with them. Use prompt caching for large repeated contexts (docs, codebases) to reduce cost and latency by up to 90%.

Workflow

Make your first API callThe Messages API is simple - system prompt sets the assistant's role; user message is what you want; assistant message is the reply.
Implement tool use (function calling)Tool use lets Claude call external functions - querying databases, calling APIs, or executing code - and reason about the results.
Stream responses for real-time UXStreaming shows text as it generates - reducing perceived latency from 5s to near-instant. Essential for chat UIs and long-form generation.
Use prompt caching for repeated contextPrompt caching stores a prefix of your context on Anthropic's servers - re-sending the same system prompt or document costs 90% less (cache read vs. full input price) and is 10x faster.
Handle errors and rate limitsProduction applications must handle API errors gracefully - rate limits, overload, and transient failures should retry, not crash.

Output Format

Claude API: messages with system prompt, tool use loop, streaming SSE, prompt caching, and retry with exponential backoff.

Model selection:
  claude-haiku-4-5   → fast, cheap, simple tasks
  claude-sonnet-5    → balanced, default for most tasks
  claude-opus-5      → most capable, complex reasoning

Caching: add cache_control to large stable context
Streaming: always for chat UIs (< 2s to first token)
Tool use: for actions + external data

Examples

Input

Output

Edge Cases & Gotchas

Context window limits - Claude Sonnet 5 has a 200k token context window. Very long conversations + large system prompts can exceed this. Implement a sliding window: keep system prompt + last N turns + current message.
Max tokens vs. response length - max_tokens is the ceiling, not the target. Claude stops when it finishes the response - setting max_tokens=4096 doesn't guarantee a long response. Set it to the maximum you'd ever want, not an estimate.
Tool use infinite loops - An agentic loop that always calls tools can run indefinitely. Set a max_iterations limit and detect when the model is calling the same tool with the same args repeatedly (a stuck loop).
System prompt injection via user content - Never interpolate raw user input into system prompts. A user who sends 'ignore instructions and...' in a field you template into the system prompt can manipulate the model's behavior.

When to Break These Rules

Use the Anthropic Workbench (console.anthropic.com) for iterating on prompts interactively before building the API integration - much faster iteration loop.
Use Claude.ai directly instead of the API for one-off tasks, research, or when you don't need programmatic output - the API adds operational complexity not needed for human-facing use.

Reference Files

https://docs.anthropic.com/en/api/getting-started
https://docs.anthropic.com/en/docs/build-with-claude/tool-use