I passed Claude Certified Architect. An exam about decisions, not definitions

I've just become one of the first certified Claude architects in Poland. And yes, I know: on LinkedIn almost everyone is an "AI architect" these days. There's one difference, though. I have the paper for it. And that paper hurt, because before you get it you sit through two hours that test your knowledge harder than I expected.

What this certification actually is

Claude Certified Architect Foundations is an Anthropic certification for solution architects. It validates one specific thing: that you can design and build production-grade applications on Claude using Claude Code, the Claude Agent SDK, the Claude API and MCP. Not "you've heard of it", but you can assemble it into a system that works at a client.

The exam is proctored and runs two hours. This is not a quiz you click through over coffee. The format treats you like a real state exam would.

Proctoring: two hours under watch

Before you even see the first question, you go through a procedure that leaves no room for cheating:

It sounds strict, and it is. But it makes sense. This certificate is supposed to mean something, so the format makes sure you can't just "click through" it.

What the questions were like

Forget memorizing the docs. This was not a "how many parameters does model X have" quiz. It was a marathon of hard case studies. Four answers, one correct, but the differences were subtle. The questions looked roughly like this:

• You have situation X. Which step do you take first?
• You have three technically correct solutions. Which is most optimal for cost and latency?

It's like choosing between two good answers in a system that actually has to run in production. One small detail in the prompting or the configuration changes everything. I used almost the full two hours, checking every question twice.

The sample questions lull you into a false sense of ease

If you're prepping from sample material, watch out for one trap. The practice questions were much easier than what I saw on the exam. It's easy to walk away thinking the target level is light. It isn't. The real exam plays a class above, so don't treat the samples as the ceiling of difficulty, treat them as a warm-up.

How to prepare

The shortest advice I have: don't learn definitions, learn to make decisions. This exam doesn't test whether you can read the docs, it tests whether you understand how Claude behaves under fire.

In practice that means drilling real trade-offs across a few areas: agentic system architecture, tool design and MCP, Claude Code configuration, prompt engineering, and context management. For each topic, don't ask "what is it", ask "when do I reach for this over that, and at what cost". Practise in pairs: cost versus latency, autonomy versus control, simplicity versus flexibility. Once you start thinking in decisions, the exam stops being a guessing game.

I later wrote up the cheat sheet I made for myself while prepping as twelve rules for building agents.

Why "decisions, not definitions" is the core of the architect's job

Because in production nobody pays you for knowing definitions. They pay you for the choice that survives contact with a real user and with the token bill. An architect is a trade-off person: someone who, out of three correct options, can point to the one that makes sense here, now, at this budget and these requirements. The exam tests exactly that muscle. And that's why its paper actually means something.

I passed well above the threshold, with a score I'm genuinely happy with. Now I'm turning that knowledge into implementations that work and create value.

One last thing. I've released the 60 questions I wrote for myself during preparation as a free practice exam in the browser: study mode, exam mode and a per-domain score. How to use it is covered in a separate post.