🔨THE BUILD
Ten lines. That's all it takes to make your first Claude API call.
But here's what most tutorials don't tell you — every single parameter in those ten lines maps to an exam domain. This isn't just a hello world. This is the foundation of everything.
Run it. Then break it intentionally. Remove max_tokens. Pass a bad model name. Drop the role key. The error messages will teach you more than the success case ever will.
stop_reason is the most exam-tested field in this entire response. After your call succeeds, you'll see end_turn. That's normal. The exam tests what you do when you see tool_use or max_tokens instead. Know those two. Cold.
📋 CERT CORNER
Domain 2: Prompt Engineering & Structured Output (20%)
Domain 2 comes down to one question the exam keeps asking in different ways.
When do you trust a prompt to get the right format? And when do you write code that enforces it?
The answer isn't complicated. If your downstream system depends on the output being structured i.e. if a database is waiting, if an API is expecting a specific shape, then you don't trust the prompt. You validate programmatically. You use Pydantic. You guarantee the structure in code, not in words.
This week's question:
A production pipeline extracts invoice data from text and writes it to a database. What's the right approach?
A. Tell Claude to always return valid JSON in the system prompt
B. Use few-shot examples showing the correct format
C. Define a Pydantic model and validate Claude's response programmatically
D. Post-process with regex to extract key fields
Why C: Prompt instructions improve compliance. They don't guarantee it. When something downstream is depending on that structure, you need a guarantee. That's what Pydantic gives you.
📦 WHAT SHIPPED
Anthropic Academy just became the most important free resource in AI certification prep. 13 courses. Mapped directly to CCA-F domains. Completely free. The Agentic AI, Prompt Engineering, and Claude Code Configuration courses are the ones to prioritize. Do those before anything else.
And on model selection, know the tiers. Haiku for high-volume tool-use. Sonnet for most production workloads. Opus where quality justifies the cost. The exam treats model selection as an architectural decision. Because it is.
🔗 THE REPO
anthropics/anthropic-cookbook
Anthropic's official code examples for every major integration pattern. If a pattern exists in production Claude systems, there's an entry here. The tool use and multi-agent sections are your Domain 1 and 4 preparation.