So I passed the Claude Certified Architect – Foundations (CCAR-F) exam — but I have to be honest with you: I failed the first time.
I want to tell you about it the way I’d tell a friend over coffee. Because here’s the thing — this exam is not easy, but it’s not scary either. It sits right in the middle. If you study the right way, you’ll be fine. Let me save you the mistake I made.
What is CCAR-F?
It’s Anthropic’s entry-level certification for people who build with Claude. The way Anthropic puts it, it’s aimed at solution architects, and it shows you can design and build production-grade applications with Claude using Claude Code, the Claude Agent SDK, the Claude API, and MCP. In plain terms: it checks that you actually understand the toolkit well enough to make good decisions.
The basics:
| Questions | 60 |
| Time limit | 120 minutes |
| To pass | 720 / 1000 |
| Price | $125 |
| Delivery | Online or at a test center |
| Valid for | 12 months |
Now, the part that trips people up — where you actually sign up:
⚠️ You can only register and pay on the official Anthropic Partner Academy certification page. That’s where you purchase the exam and book your seat — there’s no other way in.
Right now it runs through the Anthropic Partner Network, so check the sign-up details on that page before you plan around it, since that part is still changing.
One thing to know up front: it’s almost all “here’s a situation, what would you do?” questions. It’s not about memorizing facts. It’s about making the right call.
The five parts of the exam
There are five topic areas. Here’s each one, and how hard I found it:
| Domain | What it covers | Difficulty |
|---|---|---|
| D1 · Agentic Architecture | Wiring up agents and subagents | Medium |
| D2 · Tools & MCP | Designing tools, picking the right API mode | Easier |
| D3 · Claude Code Config | CLAUDE.md, rules, Skills, hooks, permissions | Hard |
| D4 · Prompt & Structured Output | Clean extraction, reliable output | Easier |
| D5 · Context & Reliability | Keeping long, multi-step work on track | Hardest |
If you’re low on time, spend most of it on D5 and D3. That’s where I struggled, and they’re the parts you can’t really fake — you have to have done the work.
Why I failed the first time
Simple: I just read everything. I went through the material, nodded along, felt ready. Then the real questions hit me — where three of the four answers all sound right — and I froze.
Reading is not studying. That was my lesson.
What I did differently for the retake
Two changes, and they made all the difference.
I turned every tricky topic into a simple rule. Instead of memorizing a definition, I wrote myself a one-liner: “if the question says this, the answer is that.” Once you know what a question is really asking, the pretty-but-wrong answers stop fooling you.
I got my hands dirty. The topics I kept missing were the practical ones — how features actually behave once you run them. So I built tiny examples: a couple of subagents, a tool or two, a workflow I could break on purpose and fix. That’s what finally made them stick.
How I studied
Nothing fancy — just 1 to 2 hours a night after work, for about 14 days. Here’s what I used:
- Official Anthropic Academy (Skilljar) courses — my main source, and the closest match to how the exam thinks. I went through a bunch: AI Capabilities and Limitations, AI Fluency: Framework & Foundations, Building with the Claude API, Claude 101, Claude Code 101, Claude Code in Action, Introduction to Agent Skills, Introduction to Model Context Protocol, Introduction to Subagents, and Model Context Protocol: Advanced Topics. If you do one thing, do these.
- A Udemy course for the theory, to make the ideas stick.
- Practice exams at CertSafari and claudecertificationguide.com — not for the score, but to find where I was weak so I could fix it.
- Real projects — I’d already been building MCP clients and servers and using Claude Code every day, so a lot of the scenarios felt familiar before I even opened a course.
A few things worth really understanding
These are the ideas I wish I’d had clear in my head the first time. Get these and a lot of the hard questions get easier:
- Parallel tool calls aren’t a setting — you ask for them. There’s no switch that turns them on. You tell the model to run independent calls together. (The one related setting only turns it off.)
CLAUDE.mdis a suggestion; hooks and permissions are guarantees. If something must always happen, don’t rely on prompting — use hooks or permissions.- Don’t force fields to be required in extraction. Required fields don’t force accuracy, they force the model to make things up. Let fields be empty and tell it to return nothing when there’s nothing there.
- When exact values matter over a long chat, keep a structured state, not a summary. Summaries drop details — that’s how the important stuff quietly gets lost.
- If a job needs to survive a crash, write progress to a file. Chat history disappears with the session. Save what’s done, and pick up from there.
Tips for exam day
- Look for the key word. Most questions turn on one word — coverage, latency, guaranteed, reversible. Find it and the answer usually follows.
- Spot the repeat offenders. The same wrong answers keep showing up (“just increase
max_tokens”, “give it every tool to be safe”). Learn them once, spot them everywhere. - Mind the clock, but don’t panic. There’s plenty of time if you keep moving — flag the hard ones and circle back.
Was it worth it?
Yeah, it was. Not really for the certificate — but because studying this way made me genuinely better at the work.
If I could tell my first-attempt self one thing: stop reading, start building. Learn the ideas well enough to recognize them, then go prove them by doing — because that’s what the exam is really measuring.
And if you fail the first time like I did — don’t sweat it. Fix how you study and go again. You’ve got this. 🙂