What to expect on the CCA-F exam (and why most candidates run short on time)

Most CCA-F preparation advice focuses on what to study. Far less of it covers what the exam is actually like to sit: how the scenarios read, where the clock goes, and why the domain that sounds easiest often isn’t.

This guide is drawn from a firsthand sitting of the exam. It covers format, pacing and the kind of thinking it tests. It contains no exam questions or recalled scenarios, and it never will.

The AI Skill Atlas verdict

CCA-F is a judgement exam disguised as a knowledge exam. Knowing the material gets you to the right two answers. Choosing between them, under time pressure, after reading a long scenario, is what decides the result.

Time is the real constraint. 60 questions in 120 minutes sounds generous. It isn’t, once scenario reading is counted.
Hands-on experience is the biggest advantage. Candidates who have built with the API, tools and MCP recognise the trade-offs. Theory-only candidates struggle most.
Don’t underestimate prompt engineering and structured output. These sound like the familiar domains. In practice they’re where close calls cluster.

Full fees, format and ratings are in the CCA-F guide.

The format, briefly

60 questions, 120 minutes, closed book. English only.
Built around 4 scenarios drawn from a pool of 6. Each scenario describes a system or business situation, and a group of questions follows it.
Online through Pearson VUE OnVUE, or at a test centre.
Passing score 720 out of 1,000. The result appears as soon as you submit.

The scenario structure matters more than any other detail. You aren’t answering 60 unrelated questions. You’re reasoning about four situations, each in depth.

Booking and check-in

Booking runs through Pearson VUE once your partner access is in place. Online proctoring through OnVUE is the convenient option, and check-in is usually smooth when you prepare for it:

Run the system test on the same machine and network a day before.
Clear your desk completely. The proctor will ask to see the room and work surface.
Log in early. Check-in takes several minutes, and the clock doesn’t start until it’s done, but arriving rushed costs focus you’ll want later.

None of this is difficult. It just shouldn’t be the first time you’ve thought about it.

The scenario format changes how you read

Each scenario is long. It sets out constraints: what the system has to do, what it can’t do, what the organisation cares about. The questions that follow test whether you noticed those constraints.

That changes the reading strategy:

Read the scenario once, slowly, for constraints. Note words like “regulated”, “low latency”, “must not”, “minimal cost”. They usually decide a question later.
Don’t reread the whole scenario for every question. Go back only to the part the question points at.
Expect questions to depend on a single detail. Two answers may both be good architecture in general. Only one fits this scenario’s constraints.

Pacing: where the clock goes

The arithmetic says two minutes per question. The reality is less, because scenario reading comes out of the same 120 minutes.

Running short on time is the most common experience, including for well-prepared candidates. A plan helps:

Budget roughly 25–28 minutes per scenario, including reading. Keep a few minutes in reserve for the end.
Check the clock at each scenario boundary, not after every question.
When two answers look right, pick the better fit, flag it, and move on. Returning later with fresh eyes works better than a three-minute debate now.
Never leave a question blank. Answer everything before you review anything.

The candidates who run out of time are rarely the ones who didn’t know the material. They’re the ones who spent too long being certain.

What the questions actually test

Three patterns stand out across the exam.

Close calls between plausible answers. Wrong options are seldom obviously wrong. Usually two are defensible, and the question turns on which better fits the stated constraints. This is where real design experience pays off.

Trade-offs over definitions. Few questions ask what a term means. Most ask which approach is better in a given situation, and why the alternative falls short. Memorised definitions help less than you’d expect.

Long setups for short questions. The work is in understanding the situation. The question itself is often one line.

The domain people underestimate

The five domains are agentic architecture, MCP integration, prompt engineering, context management and structured output.

Agentic architecture and MCP get most of the preparation attention because they sound advanced. Prompt engineering and structured output are the ones that catch people out. Most candidates use prompts and JSON output every day, so they assume these domains are free marks.

They aren’t. Questions here tend to compare two approaches that would both work, and ask which is more reliable, more maintainable or better suited to the scenario. Everyday familiarity breeds confidence; the exam rewards precision. Give these domains as much deliberate practice as the “hard” ones.

How to prepare

A preparation window of one to two months is realistic for someone with hands-on LLM experience. Less suits people who already build with Claude daily; more suits those coming from other platforms.

What works best, roughly in order of value:

Hands-on building. Build something with tool use, an MCP integration and structured output. The trade-offs the exam asks about become obvious once you’ve hit them yourself.
Anthropic’s documentation and free courses at anthropic.skilljar.com. These define the vocabulary and the recommended patterns the exam assumes.
Real project experience, if you have it. Any system you’ve designed, debugged or explained to a stakeholder is preparation.
Timed scenario practice. Practise under time pressure, not just for accuracy. Reading speed and decisiveness are skills that improve with repetition.

What to avoid: memorising definitions in isolation, and practising only short standalone questions. Neither prepares you for long scenarios against the clock.

Who will find it hard

Theory-only learners struggle most. People who have read widely about LLMs but never built anything find the close calls hardest, because they’ve never felt the consequences of choosing the wrong approach.

If that describes you, spend a few weeks building before you book. The who should NOT take a Claude certification guide has suggestions for what to build.

After you submit

The result appears on screen straight away. The credential is valid for 12 months, renewal is free, and a digital badge follows. If you don’t pass, the retake costs the full fee again, with a minimum 14-day wait before a second attempt.

Is it worth it?

For someone at a partner company who designs or leads AI systems, yes. CCA-F is one of the more practical AI certifications available, precisely because it tests judgement rather than recall.

If you’re not sure CCA-F is the right first exam for your role, start with Which Claude certification should you take first?

Full details on each exam are in the guides linked above, and you can see all of them side by side on the comparison page. If you’re preparing for an AI certification exam, try our free practice material and prep kits.