Mit Claude auf die Claude Certified Architect Foundations Prüfung vorbereiten

I've never built an agent before, not a multi-agent system, not even a single agent with a tool loop but preparing for and passing this exam gave me a better understanding of how to architect solutions with Claude.

I did most of my preparation with Claude itself which feels a bit like asking the student to mark their own homework. It turned out to be a really good study partner with a few caveats that I'll get to. As always a certification is no match for practical experience. This post covers what I did, what worked and what didn’t.

I decided to go for this certification and started preparing end of August. I fed these prompts into Claude and got it to explain things to me. I'm a sucker for handwritten notes so I went through all the domains and wrote down the key bits, comparing them against the exam guide to make sure I didn't miss anything.

Life and work got busy as usual and my prep went on and off. I'd pick it up for a few days, get pulled away and drop it again. Once I'd covered all the domains and done some practice questions I got sick of just preparing so decided to book the exam to sit in 2 days time.

In the run up to the exam, I worked through the material with Claude again then did a few rounds of practice question with a short mixed round written by Claude and some community question sets I found online.

The exam wasn't too bad and I felt fairly calm going into it. There were some questions that had me questioning whether I'd done enough prep or not but once I got my results all of these worries faded. I ended up with 802 against a pass mark of 720.

I gave Claude the study material for each domain and asked it to teach me as someone who had basic knowledge and exposure to Claude. It went through one task statement at a time explaining the concept with a production example, flagging the exam traps, asking me a check question and then connecting it to the next topic. Having to answer in my own words before being told whether I was right helped to solidify my understanding.

At the end of each domain it gave me a practice exam with one question at a time. I'd reason through each before seeing the answer and I later asked it to shuffle the order and hide the domain labels so I couldn't predict what was coming. This felt much closer to the real thing.

Once I'd been through all five domains, I tested myself on two practice sets I found online: OlivierAlter's question set on GitHub and Matthew Purcell's question set. Although both are unofficial, they were still quite an accurate representation of the actual exam. Claude presented the questions, I answered one at a time and it explained each of the options even the wrong ones and why they were wrong.

Claude was able to find patterns in the answers I repeatedly got wrong and this was the most useful part. My practice scores were fine but what mattered was where and why I lost the points. It wasn’t that I didn't know the concept, it was that I was reacting to a familiar cue instead of the detail that decided the question. Seeing the same slips come up across domains made them easy to fix on the next round.

• Being made to answer first then having every wrong option explained, not only the right one • One question at a time shuffled and with no domain labels so I couldn't coast on knowing which topic was coming • Getting walked through the reasoning when I was stuck. On one question I said I didn't know and Claude showed me how to work it out step by step instead of just giving me the answer • A short list of rules that I could skim before the exam, for example an empty result with no error flag is a valid answer, not a failure

• My practice scores sometimes lulled me into a false sense of security. I got feedback after every answer and saw the same ideas repeated across sets whereas the real exam gives you no feedback at all. Don't read your practice percentage as your exam score. • Claude and the practice sets weren't always right. At one point I pushed back on how Claude had marked one of my answers, it first defended itself with a plausible sounding explanation then admitted it had overcorrected. Treat every answer as a starting point, Claude's included, and check anything against the official docs.

• Get asked the question before you see the answer. Reading an explanation feels like learning but answering is what shows you what you don't know. Make sure to focus on details that separate what’s being asked from the pattern you recognise • Look at why you got a question wrong, not only what the answer was.

My misses came down to a handful of repeated slips so it’s good to constantly review your notes • Verify what Claude tells you against the source and expect it to be wrong occasionally • Be consistent with your prep wherever you can. A few days here and there across two months meant I could still take my time preparing for the exam rather than cramming everything. • Get hands on if you can. Practice questions only get you so far when it comes to working with AI agents in production.

I didn't do much hands on practice for this exam so that would be mainly where my gaps still lie. The exam gave me a good understanding of how I can get started with AI agents in production but it would of course be different to actually working with them in real life.

One tip for exam day: if you sit the exam online with Pearson Vue, you can strike out the answers you think are wrong whilst working through the options. It helped to declutter my screen so I could focus on reasoning through the ones that seemed correct.

If you’re sitting this exam and have any questions about how I prepared, feel free to reach out in the comments. Good luck!