“The practice questions helped me strengthen my understanding of Claude architecture and approach scenario-based questions with greater confidence.”
Niharika S
AI Solutions Architect
CCAR-F doesn’t test recall the way most IT certifications do. You won’t see “what does MCP stand for” style questions. Instead, Anthropic builds the exam around six named production scenarios — things like a customer support resolution agent, a multi-agent research system, or wiring Claude Code into a CI/CD pipeline — and draws four of the six for any given sitting. Every question you answer is attached to one of those four scenarios, so you’re reasoning about a specific, system rather than answering in a vacuum.
That structure changes how you should prepare. You get 60 multiple-choice and multiple-response questions in 120 minutes, which works out to two minutes per question — tight enough that you can’t afford to re-read a scenario three times before you understand what’s being asked. A scaled score of 720 out of 1,000 is required to pass, and because you don’t know in advance which four of the six scenarios you’ll get, there’s no shortcut to skipping any of them in prep
The difficulty pattern is consistent: most wrong answers aren’t obviously wrong. They’re plausible-sounding fixes that solve the wrong problem, or solve the right problem with more machinery than it needs. That’s the muscle this practice platform is built to train.
Domain | Weight | What It Tests |
|---|---|---|
Agentic Architecture & Orchestration | 27% | Agentic loop lifecycle, orchestration patterns, guardrails, the Claude Agent SDK |
Claude Code Configuration & Workflows | 20% | Settings hierarchy, CLAUDE.md scoping, hooks, permissions, CI/CD integration |
Prompt Engineering & Structured Output | 20% | System prompt design, structured output validation, few-shot use (and misuse) |
Tool Design & MCP Integration | 18% | Tool schema and description design, MCP servers, clients, and routing |
Context Management & Reliability | 15% | Context windows, prompt caching, long-conversation handling, production reliability |
Domain 1 alone accounts for more than a quarter of your score. Combined with Domain 3, that’s 47% of the exam sitting in just two domains — worth knowing if your study time is limited.
A candidate who studies all five domains evenly is already behind someone who weighted their prep to match the exam. Fifteen percent on Context Management doesn’t mean it’s safe to skip — it means roughly nine questions, which is still enough to fail you if you get them all wrong. But it does mean Domain 1 deserves disproportionate attention first.
Detecting when the agentic loop is actually finished. This trips up more people than any other single concept on the exam. The instinct is to check whether Claude’s response looks like plain text, or to scan the output for phrases like “I’m done.” Both are wrong. Claude can return text alongside a tool-use block in the same response, so text appearing doesn’t mean the task is complete. The only reliable signal is the stop_reason field — specifically, waiting for end_turn. Candidates who rely on hardcoded iteration caps as their primary exit condition also get caught out, because a fixed cap either kills a task that needed a few more cycles or burns cycles the model never needed.
Tool misrouting and what actually fixes it. When a scenario shows Claude calling the wrong tool, the instinctive answer is to add a routing layer or consolidate tools. The exam consistently rewards the simpler diagnosis: Claude selects tools primarily by reading their descriptions, not their names or parameter schemas, so a vague description (“Retrieves customer information”) is almost always the cause.
The CLAUDE.md scoping exception. Most candidates learn the settings hierarchy fine — enterprise overrides user, user overrides project — and then miss the exception that CLAUDE.md files at the working directory level are always included regardless of that hierarchy. It’s a small distinction, but the exam uses it to separate candidates who memorized the rule from those who understand how it actually behaves.
Progressive summarization silently destroying accuracy. When conversation history gets compressed to save tokens, specific values — a dollar amount, an order number — often get flattened into vague references like “a recent order.” The model doesn’t flag this; it just answers confidently with the wrong number. It’s a low-weight domain (15%) but a disproportionately common exam theme.
Because every question is scenario-anchored, drilling isolated flashcard-style facts won’t get you far here. The more useful habit is reading a full scenario, generating your own answer before looking at the options, and only then comparing your reasoning against the four choices. CCAR-F is deliberately built so that two or three answers sound reasonable — the exam is testing whether you can identify the one that solves the stated problem without over-engineering it.
Pay attention to explanations, not just whether you got a question right. The “simplest fix that works” pattern shows up across Domains 1, 2, and 4 in different disguises, and the only way to internalize it is to see it explained repeatedly across different scenarios until you start predicting it before you read the answer choices Begin your preparation with a free VMware Cloud Foundation 5.2 Administrator practice test.
Since you won’t know which four of the six core scenarios you’ll draw on exam day, resist the temptation to specialize. Practicing heavily on multi-agent research systems while skimming past CI/CD scenarios leaves you exposed if that’s one of your four.
Studying all five domains with equal time instead of weighting toward Domains 1 and 3, which together make up nearly half the exam.
Memorizing API method signatures instead of understanding what fields like stop_reason actually mean and when to check them.
Only learning the correct way to build something, when roughly half the exam asks what’s wrong with a flawed approach.
Treating few-shot examples or additional model layers as default fixes, when the exam frequently rewards a narrower, cheaper fix like a clearer description or more explicit, measurable criteria.
Skipping scenario practice entirely and relying on domain summaries — the exam’s difficulty lives in distinguishing between plausible answers, which summaries alone don’t train you for.
Already building production Claude systems (6+ months hands-on with the APIs, Agent SDK, Claude Code, and MCP): One to two weeks of focused review, concentrated on the anti-patterns in Domain 1 and the CLAUDE.md scoping exception in Domain 3, should be enough alongside a full-length practice run.
Comfortable with Claude’s API but limited agentic/MCP experience: Budget three weeks. Spend the first week on Domains 1 and 3 since they carry the most weight, the second on Domains 4 and 2, and the final week on Domain 5 plus at least one full-length timed practice exam.
Retaking after a previous attempt: Go back through your scored practice attempts domain by domain rather than restarting broadly. Most retake candidates fail on a narrow set of recurring traps — loop termination logic, tool description design, or the progressive summarization failure mode — not on broad unfamiliarity with the material.
This is worth flagging before you invest study time: CCAR-F registration currently requires an email domain tied to an organization in the Claude Partner Network, and Anthropic’s Partner Network entry criteria have included a minimum commitment of ten certified practitioners for new applicants. Confirm your access through your organization or Anthropic before committing to a study timeline — it’s a detail some candidates discover only after they’ve already started preparing.
Last updated on Nov, 16 2026