Using the Claude API in Workflows

While pre-built interfaces and CLI agents (like Claude Code) are excellent for standard tasks, building custom AI workflows requires interacting directly with the Anthropic Claude API.

Direct API integration allows you to automate code reviews, build custom repository indexing scripts, orchestrate multi-agent runs, and steer model behavior with high precision.

---

1. Setup & SDK Installation

Anthropic provides official libraries for Node.js and Python. We recommend using uv to run scripts instantly with their dependencies without polluting your global environment.

Python Setup (using uv)

Run your script directly with the anthropic library dependency declared:

uv run --with anthropic python script.py

Authentication

Set the API key as an environment variable in your terminal or .env file:

export ANTHROPIC_API_KEY="your-api-key-here"

---

2. Core API Implementation

The primary entry point for Claude is the Messages API. Here is a production-ready Python example demonstrating system prompt steering and the messages array:

import os
from anthropic import Anthropic

# Initialize client (looks for ANTHROPIC_API_KEY in environment)
client = Anthropic()

response = client.messages.create(
    model="claude-3-5-sonnet-latest",
    max_tokens=2000,
    temperature=0.2,
    system="You are an expert Incrementic coding agent. Always output clean, vanilla CSS and follow plain over clever formatting.",
    messages=[
        {
            "role": "user",
            "content": "Create a CSS design token for our primary brand color (Incrementic Red, #EA5148)."
        }
    ]
)

print(response.content[0].text)

---

3. Advanced Workflow Patterns

To build fast, economical, and robust workflows, you must leverage Claude-specific API features:

A. Prompt Caching (Essential for Codebases)

When building coding agents, you frequently send large files or repo contexts in your prompts. Claude's Prompt Caching allows you to cache these static contexts (like system instructions or base codebases) to reduce costs by 85% and latency by 90%.

[!IMPORTANT] To enable caching, add the cache_control block to the end of the static content you want to cache (e.g. system instructions or a large file block):

{
  "role": "user",
  "content": [
    {
      "type": "text",
      "text": "... 50,000 lines of codebase files ...",
      "cache_control": {"type": "ephemeral"}
    },
    {
      "type": "text",
      "text": "Find where the primary color is defined and replace it."
    }
  ]
}

B. Steering with XML Tags

Claude is specifically trained to structure its thinking, context, and responses using XML tags. Always wrap dynamic inputs in clean tags to guide the model's focus.

• Prompt Pattern:

  You are a code refactoring assistant.
  
  <codebase_rules>
  - Always use spaces, not tabs.
  - Primary color is #EA5148.
  </codebase_rules>
  
  Please refactor the following source code:
  <source_code>
  const color = "green";
  </source_code>

C. Tool Use (Function Calling)

You can equip Claude with custom "tools" (like reading a file, running a shell command, or calling a search API). The model will decide when to call these tools based on the user request.

response = client.messages.create(
    model="claude-3-5-sonnet-latest",
    max_tokens=1024,
    tools=[
        {
            "name": "read_file",
            "description": "Read the contents of a file in the workspace",
            "input_schema": {
                "type": "object",
                "properties": {
                    "path": {"type": "string", "description": "Absolute path to the file"}
                },
                "required": ["path"]
            }
        }
    ],
    messages=[{"role": "user", "content": "Read /src/css/custom.css"}]
)

# If Claude decides to call the tool, response.stop_reason will be "tool_use"

---

Best Practices

• Temperature Control: Use low temperatures (0.0 to 0.2) for deterministic coding tasks. Use higher temperatures (0.5 to 0.7) for brainstorming, naming products, or writing blog posts.
• System Prompt Isolation: Put style guidelines, allowed tools, and rules in the system parameter rather than mixing them into the messages array.
• Handle Rate Limits: Build retry mechanisms (like exponential backoff) into your custom API workflows to handle rate limits gracefully during large concurrent operations.