The mistake I would avoid when preparing for the Claude Certified Architect Foundations exam is treating it like a vocabulary test.

Vocabulary helps. You need to know what MCP is, what a tool schema does, why context management matters, and how Claude Code changes the operating surface. But the exam is much closer to an architecture review than a flashcard deck.

It gives you a system in trouble and asks what you would do next.

The useful study question

When you read a scenario, do not start by hunting for the familiar term. Start with the failure.

Ask:

  • What is the agent trying to do?
  • Where did the loop, tool, context, or permission boundary break?
  • Which answer reduces ambiguity instead of adding more automation?
  • Which answer leaves a better trail for review and recovery?
  • Which answer sounds productive but widens the blast radius?

That last question is where many wrong answers hide.

Prompt fixes are often too small

A lot of agent failures look like prompt problems at first glance. Sometimes they are. Often they are architecture problems wearing a prompt-shaped mask.

If a tool returns “failed” with no structured error, a better prompt does not give the agent a recovery path. If a subagent is missing the source facts, a clearer instruction does not create context. If Claude Code has permission to edit too much of the repo, a polite warning is not a control.

The exam tends to reward the answer that changes the system boundary, not the answer that merely asks the model to behave better.

Study the wrong answers

For CCA-F prep, I would spend less time asking “what is the right answer?” and more time asking “why are the other three wrong?”

That is uncomfortable, but it works. The wrong answers teach the principle.

One answer may ignore human review. Another may make the tool contract vague. Another may centralise too much context in the wrong place. Another may automate a step that should have an approval gate.

Once you can name the violation, the right answer becomes easier to see.

The exam skill is also the job skill

This is why I like the shape of the exam. It rewards the same instincts that matter in real Claude architecture work.

Bound the agent. Give tools clear contracts. Keep context deliberate. Make retries finite. Preserve evidence. Escalate when the system no longer has enough information to act safely.

That is good exam prep, and it is good engineering.

I wrote Architect the Agent: The CCA-F Certification Guide around that premise. The book is now live on Leanpub: get the guide here.