What the exam is
The Claude Certified Architect, Foundations exam is Anthropic's credential for people who design and build production systems on Claude. It is proctored, closed book, and scored on a scale where 720 out of 1000 passes. Every question is a scenario: a support agent that must decide when to escalate, a tool that fails silently in production, a teammate whose Claude Code sessions ignore the project conventions. You are asked which design is right and, implicitly, why the three plausible alternatives are wrong.
The published blueprint weights five domains. Agentic architecture and orchestration is the largest at 27 percent, followed by Claude Code configuration at 20, prompt engineering and structured output at 20, tool design and MCP integration at 18, and context management and reliability at 15. Candidates who have taken it consistently say the agentic domain is the hardest, because it rewards architectural judgment over remembered syntax.
Why the usual study plan is slow
The default approach is to read the documentation, watch a course, and take a practice exam at the end. It feels productive and it is mostly wasted. Reading produces recognition, not recall. You finish a page on prompt caching feeling like you understand it, and a week later you cannot say which change to a request invalidates the cache. The exam asks for recall under pressure, and the only way to build that is to practice retrieving the answer, get corrected immediately when you are wrong, and come back to that exact point before you forget it again.
Three findings from learning research do the heavy lifting here. Retrieval practice: answering from memory strengthens the memory far more than re-reading. Immediate feedback: a correction that arrives seconds after the error sticks; one that arrives at the end of a 60-question test barely registers. Spaced repetition: revisiting an item at growing intervals, a day, three days, a week, moves it into long-term memory with a fraction of the total time.
There is a fourth that most study tools skip. If you were confident and wrong, that correction lands harder than any other kind. Researchers call it the hypercorrection effect. The practical implication is that you should rate your confidence before you see the answer, so the tool knows which misses to lean on.
What we built
The trainer applies all four. Each round is ten scenario questions, mixed across domains the way the real exam mixes them. You pick an answer and rate how sure you are. Then you see whether you were right, the reasoning behind the correct option, and one line on why each wrong option is wrong. A miss comes back a few questions later in the same round, and again the next day. A correct answer you guessed at does not count as mastered. Questions you have answered confidently on several separate days stop appearing, so your time goes to what you do not know yet.
Mock exams draw 30 questions in the official domain proportions, favoring whatever you have not mastered. A perfect mock therefore means something. Everything is stored in your browser; nothing is sent anywhere.
One thing worth saying about the questions themselves. The first draft made a classic mistake: the correct answer was always the longest and most technical option. Anyone who has taken a few certification exams can spot that pattern without knowing the material. Every option is now the same length and the same register, and the wrong answers are real techniques applied in the wrong place. A PreToolUse hook where a PostToolUse hook belongs. Prompt caching where you actually need persistent memory. You have to know why, not just recognize the jargon.
How to use it
- One or two rounds a day. Ten questions takes about ten minutes. Daily practice with spacing beats a weekend marathon every time.
- Read the corrections on questions you got right. The one-line reasons why the other options fail are where the near-miss knowledge lives.
- Take a mock every few days. It is the only honest measure of whether you are ready. Aim for two consecutive mocks above 90 percent before booking.
- Build the things. No question bank replaces having written an agent loop, wired a hook, or watched a cache miss in a real request. The trainer tells you what you do not know. The docs and a terminal fix it.
A note on what this is not
This is an unofficial study aid written from the public blueprint and the documentation. It is not affiliated with Anthropic, and no question comes from the real exam. Exam format, pricing and registration details change; check Anthropic's certification pages before you book.