Lessons from the Anthropic Architect exam
CCA-F facts and falacies.
The Claude Certified Architect Foundations exam (CCA-F) was the only official certification from Anthropic until July. Having taken it recently I thought I'd dispel some rumours, share my approach, and review my mistakes for anyone thinking of sitting it.
Headline takeaway: more trad coding than AI research
The bulk of the curriculum focuses on wrapping the models in deterministic gates with human escalation. Scaffolding the models in code and tooling for reliable results, observability and governance.
So much for AI replacing coders!
Key facts
The certification was announced some time before its release, creating an information vacuum promptly filled with internet guesswork. Here’s the reality:
Certification lasts a year.
The cooling-off period for resits starts at 14 days, not six months (it stretches to 30 and then 90 days if you keep failing, with four attempts allowed per rolling year).
The cert is still gated behind the Claude Partner Network (CPN). Release outside the CPN is planned but not scheduled.
You do not need ten people in your organisation to pass the CPN induction before you can register.
You do need ten certified people to progress to the next partner tier.
The list price is $125, but CPN discounts apply, seemingly on a whim.
The exam is 60 questions in 2 hours.
Scores are scaled out of 1,000 with a pass mark of 720, so it's not a straight percentage.
The official practice exam is no longer available, which is probably a good thing, as I'll explain below.
Exam conditions
You can choose between a physical test centre and an online proctored exam. I chose the latter. Pearson VUE do the proctoring and it's worth reading the instructions well in advance. There are tight controls around VPNs and proxies (you can't have any), and your workspace must be clear with a single monitor. If you have two monitors, you must physically disconnect one.
All of that is checked by a downloadable app to run in the waiting room on exam day. You can also run it as soon when you schedule the exam. Don't skip that.
Be warned, the system config check is not enough. The invigilator will want to see the loose monitor cable. My webcam is fixed to my screen, or at least it was. Pointing it at a disconnected cable, itself fixed and hidden beside the dock under my desk, caused significant hassle.
Oh, and don't wear a watch. Not just smartwatches; even clockwork bling is banned.
Prep
I over-prepared, or so I thought. I passed, but not by a large margin. The exam questions are denser and more tricksy than either the official practice exam or training material.
The official mock gave a false sense of how you'll perform on the day: I passed the mock cold, in plenty of time and by a good margin. In the real thing I had ten minutes to spare and questions to revisit.
The clarity of heindsight
I scored poorly on my strongest areas and very well on my weaker ones. There were two reasons for this. Firstly, I assumed regular use would carry me through that part of the curriculum. On reflection, that was a mistake.
I use a number of AI models and harnesses that have similar concepts and subtly different configuration philosophies and permissions. It's easy to confuse Cursor and Codex conventions with Claude's.
In day-to-day work I rely on tools that abstract the underlying mechanics: a bespoke hub that distributes and security gates skills, hooks and MCPs across harnesses. I also use code discovery tools like CodeGraph that cut down on tokens and plumbing. It’s been a while since I worked wthout those and there were questions on their Claude native equivilents.
Second, I leaned too heavily on Claude-generated drills. Not only were they biased towards certain areas, they produced questions orders of magnitude easier than the actual exam. The real questions are dense; it takes time to identify the concept being tested and dodge the distractors.
A better approach would have been to stick to my original plan:
Create a side project that encapsulates cert. scope, with scaffolding so you don't get sidetracked on non-AI plumbing. Use placeholders and TODOs to show where you hand-code the exam content. Yes, trad coding. I created an AI governance control plane repo that I regret not finishing.
Go heavy on labs across the full curriculum. Python notebooks are good here.
Work up a Claude skill that generates drill questions as dense and tricky as possible from the course documentation. Fable does a much better job of question generation than Opus, at least until your drill skill matures.
Exam dumps and prep sites are worth a cursory look, but many questions are questionable and a some are just plain wrong.
The certs
Anthropic extended their cert suite in early July, with prep content arriving on 13th July, again through the CPN and Anthropic Partner Academy :
Claude Certified Associate – Foundations ($99), for consultants, sellers and delivery leads who use Claude in their own work
Claude Certified Developer – Foundations ($125), for engineers building with the Claude API, Claude Code and MCP
Claude Certified Architect – Professional ($175), for architects leading enterprise-scale deployments
The original Architect – Foundations stays in the line-up, repriced at $125.
Takeaways
This isn't a cram-and-slam exam. That's a good thing; it gives the cert value.
The more you use the full set of core Anthropic tools daily, the easier you'll find it.
Don't assume six months of sophisticated use will get you a pass without at least some prep. Unless you're Boris Cherny.
Certifications are ridiculed by many aficionados who claim they produce qualified but inexperienced practitioners. True, but I've generally considered that view too narrow. Certification forces coverage that day-to-day experience quietly skips. That said, a cert should be desirable on a job spec, never a requirement.
I know there are those who have passed with no prep, in half the allotted time, with a score of 93%. Good for you. For most of us it's worth actually working for, and that's what makes it worth having.
I’ll be taking the Professional Architect cert next. Partly because I see value in these qualifications, but mostly out of indignation at a less-than-perfect score on the previous one.