How Hard Is the CCA Exam? Judgment, Not Memory

The CCA Foundations (CCA-F) exam is moderately hard, and hard in a specific way: it tests engineering judgment about Claude, agents, tools, and context rather than recall of documentation. Sixty scenario-based questions in 120 minutes, closed book, with a scaled pass mark of 720 out of 1000. Confirm those figures on the official Anthropic Academy or Pearson VUE page before you register.

If you can already explain why a design fails, not just what the docs say, the exam is passable with a few weeks of targeted review. If you have only read about Claude, expect a steeper climb. This post breaks down what kind of hard it is, which domains cause the most trouble, and what makes it easier.

What kind of hard is the CCA-F exam?

Not the memorisation kind. Very few items ask you to recall a parameter name in isolation. Instead, a typical stem describes a working system, introduces a constraint (budget, latency, coverage, isolation, a rule that a human must approve something), and asks which change is the least bad. All four options are usually things a competent team might try; one of them respects every constraint and the others each quietly break one.

That structure produces two kinds of difficulty:

• Recognition under noise. You must spot the underlying mechanism (tool-count degradation, lost-in-the-middle attention, stateless API history, prompt vs programmatic enforcement) inside a story about a support desk or a research pipeline.
• Trade-off discipline. You must reject the “best-sounding” option because it violates the stated constraint, even though it would be great without it.

People who read docs but rarely build tend to struggle with the second kind. People who build but skip the reading struggle with the first.

Which domains trip candidates up most?

Based on the domain weights and the kind of judgment each domain demands, here is where the pain usually is. Anthropic has not published per-domain difficulty data; this is our read of the blueprint and of how practice-question performance clusters.

Domain	Weight	Why it is hard
D1 Agentic Architecture & Orchestration	27%	Largest domain, longest scenarios: sub-agents, context passing, hooks, resume vs fresh session. Many candidates know one pattern (say, hub-and-spoke) and mis-apply it everywhere.
D5 Context Management & Reliability	15%	Fewest questions but the subtlest trade-offs: when to escalate, whether a confidence score means anything, how to report a failed search without fabricating a conclusion. Two options often look equally sane.
D2 Tool Design & MCP Integration	18%	Easy if you have written tool descriptions and watched routing fail; abstract if you have not. Error-classification questions are a common miss.
D4 Prompt Engineering & Structured Output	20%	Familiar territory for most, but the exam pushes past “write a good prompt” into few-shot design, nullable fields, and validation-with-feedback loops.
D3 Claude Code Configuration & Workflows	20%	Straightforward for daily Claude Code users; genuinely hard for anyone who has never touched CLAUDE.md, skills, hooks, or path-scoped rules.

Two observations. First, D1 is where the exam is won or lost simply because there are roughly 16 questions in it. Second, D5 punishes overconfidence: it is the domain most likely to make you pick a “reasonable” answer that ignores what the scenario said about the customer, the outage, or the missing data.

What makes the exam easier?

• Hands-on time. An afternoon writing a project CLAUDE.md with path-scoped rules, or wiring an MCP server and watching a tool get misrouted, converts abstract items into recognisable ones. See the Domain 3 and Domain 2 guides for what to build.
• Reading the constraint first. Before you look at options, find the sentence that begins with “must,” “cannot,” or “without.” Eliminate any option that violates it. This alone reclaims several questions and 15 minutes.
• A domain-first study order. Cover D1 thoroughly, then D3 and D4, then D2, then D5. Weight-adjusted, that ordering gives you the most points per hour. Details in our domain weighting post.
• Explanations, not just answers. After every practice question, say why each wrong option is wrong. The practice hub has 1,374 questions with explanations for exactly this.
• Timed mocks. Our 60-question mock uses a 90-minute clock, tighter than the real 120, so exam-day pace feels easy by comparison.

What is the exam really testing?

A short worked illustration. Suppose a scenario says a research agent’s search tool times out during an outage and returns an empty list, and asks how the final report should treat that topic. The recall-level answer is “handle the error.” The exam-level answer distinguishes an access failure (we could not look) from a valid empty result (we looked and found nothing) and expects the report to annotate coverage rather than conclude that nothing exists. That is a judgment about reliability, and it is covered in Access failure vs valid empty. Nearly every question has that shape: a mechanism you can learn, wrapped in a story you have to read carefully.

Is it passable on the first attempt?

Yes, for candidates who do three things: build a little with the tools, study by domain weight, and take timed mocks until they clear 80% twice. Anthropic has not published pass rates, and salary or job-market claims tied to the exam are not verifiable, so ignore hype in both directions. What you can measure is your own mock trend, and it predicts the outcome better than any forum thread. If you are far from 720-equivalent today, that is a plan, not a verdict.

Next step

Take the free CCA-F mock exam cold to see which kind of hard hits you: recognition or trade-off. Then start with the Domain 1 study guide; it is the largest domain and the one where the exam’s least-bad-design framing is most concentrated.