Claude Certification Sample Questions: CCAR-F Exam Guide

Every candidate asks for claude certification sample questions before sitting the CCAR-F, and for good reason: the exam tests applied judgement across five domains, not vocabulary. Below we work through the question format, provide annotated examples from each high-weight domain, and explain the reasoning pattern the exam consistently rewards.

What format do CCAR-F sample questions take?

Each of the 60 items presents a paragraph-length scenario and asks you to select the best course of action, the root cause of a failure, or the correct architectural choice. Some items are multiple-response: you must select exactly the number of correct options stated in the stem. The 120-minute time limit gives you roughly two minutes per question.

Per Anthropic's exam guide, each sitting draws four scenarios at random from a bank of six. That randomness makes broad preparation essential: you cannot rely on a predictable domain distribution, and skipping any domain is a genuine risk.

Which domains should I front-load when studying?

The five CCAR-F domains and their exam weights are:

Domains 1, 3, and 4 together account for 67% of the exam. If your study time is limited, prioritise these three first. Domain 1 at 27% alone justifies at least a third of your preparation hours.

Our concept library at /concepts maps 174 atomic concepts to these five domains and 30 task statements, so you can identify gaps precisely rather than studying by instinct.

What does a Domain 1 sample question look like?

Scenario. An orchestrator spawns three research subagents in parallel. Two return results promptly, but the third never sends a response. The orchestrator's loop exits after the two successful results, and the final synthesis is incomplete. A junior engineer suggests adding a longer timeout. What should the architect do first?

A. Increase the timeout parameter on the subagent call. B. Inspect the stop_reason field on the missing subagent's last message. C. Route the failed subagent's work to a fallback agent immediately. D. Replace the parallel pattern with a sequential pipeline.

Correct answer: B.

The exam rewards root-cause tracing before remediation. The stop_reason field tells you whether the subagent completed, hit a token limit, or stalled on a tool call. Increasing a timeout (A) treats a symptom without diagnosis. Routing to a fallback (C) is a valid recovery step, but only after you know what failed. Replacing parallel with sequential (D) is a disproportionate change that sacrifices throughput for a problem whose cause is still unknown.

This pattern recurs across agentic architecture questions: the exam penalises disproportionate fixes. Adding complexity without diagnosing the root cause is consistently the wrong answer.

What does a Domain 2 sample question look like?

Scenario. An agent has access to two tools: search_documents and fetch_document. The descriptions read: search_documents returns "documents"; fetch_document "gets a document." The agent repeatedly calls fetch_document with natural-language queries instead of structured IDs, causing errors. What is the most effective single fix?

A. Add a system prompt instruction telling the agent which tool to call for search queries. B. Rewrite both tool descriptions to distinguish query-based search from ID-based retrieval. C. Remove fetch_document from the agent's tool list. D. Add input validation to fetch_document that rejects non-ID strings.

Correct answer: B.

Tool descriptions are the primary selection mechanism the model uses when deciding which tool to invoke. Vague descriptions cause misrouting. Rewriting them is a low-effort, high-leverage fix that corrects the root cause. A system prompt instruction (A) adds a second layer of guidance but leaves the broken description in place, which is less reliable and harder to maintain over time. Removing the tool (C) destroys capability. Input validation (D) prevents errors at call time but does not fix the selection problem.

A corrected description pair looks like this:

{
  "tools": [
    {
      "name": "search_documents",
      "description": "Full-text search across the document corpus. Accepts a natural-language query string. Returns a ranked list of document IDs and excerpts."
    },
    {
      "name": "fetch_document",
      "description": "Retrieves the full content of a single document. Requires an exact document ID (e.g. 'doc_1234'). Do not use for search or exploration."
    }
  ]
}

Domain 2 questions frequently test the principle that diagnosing tool misrouting starts with the description layer, not the system prompt.

What does a Domain 3 sample question look like?

Scenario. A team runs Claude Code in a headless CI pipeline using a shared project settings.json. A developer wants to allow a specific file-write operation in their local environment without affecting CI or other team members. What is the correct approach?

A. Edit the shared settings.json to add the permission globally. B. Create a settings.local.json in the project directory with the permission. C. Add an environment variable override in the CI configuration. D. Use a custom slash command to bypass the shared settings.

Correct answer: B.

The three-level configuration hierarchy in Claude Code covers user, project, and local levels. settings.local.json is gitignored by default, so local overrides do not propagate to CI or teammates. Editing the shared file (A) changes behaviour for everyone on the team. Environment variables (C) are not the correct mechanism for permission scoping. A custom slash command (D) does not interact with the permission layer at all.

{
  "permissions": {
    "allow": ["Write(src/output/*.json)"]
  }
}

The version control implications of which configuration files are committed versus gitignored are a recurring source of exam items in Domain 3. Understanding which settings travel with the repository and which stay local is tested directly.

What does a Domain 4 sample question look like?

Scenario. A pipeline extracts product attributes from unstructured text and returns JSON. In production, the model occasionally omits required fields or adds fields outside the schema. An engineer proposes adding a system prompt note: "Always fill every required field." What is the more reliable solution?

A. Add the system prompt instruction and monitor for improvement. B. Enforce output shape using a JSON schema with required fields and additionalProperties: false. C. Post-process responses with a validator that inserts null for missing fields. D. Switch to a larger model.

Correct answer: B.

Prompt Engineering & Structured Output questions consistently favour schema-level enforcement over instruction-level enforcement. A system prompt note (A) is probabilistic; the model may still deviate under distribution shift. A validator that inserts null (C) silently masks production data-quality failures and obscures the underlying problem. Switching models (D) may improve compliance but does not guarantee it and increases cost unnecessarily. A strict JSON schema constrains the output space deterministically:

schema = {
    "type": "object",
    "properties": {
        "product_name": {"type": "string"},
        "price_usd":    {"type": "number"},
        "category":     {"type": "string"}
    },
    "required": ["product_name", "price_usd", "category"],
    "additionalProperties": False
}

The exam returns to this theme across all five domains: when the cost of failure is high, prefer a deterministic constraint over a probabilistic instruction.

What does a Domain 5 sample question look like?

Scenario. A research pipeline processes 40 documents in a single session. Early in the session the agent accurately attributes findings to source documents. By document 35, attribution has degraded: the agent conflates sources and omits provenance details. The engineer considers increasing the model's context window. What is the correct diagnosis?

A. The model has a defect that worsens over long sessions. B. The pipeline is experiencing attention dilution across the extended context. C. The context window is too small to hold all 40 documents. D. The summarisation step is removing attribution data prematurely.

Correct answer: B.

Context Management & Reliability questions frequently test the attention dilution problem: as context grows, the model's ability to attend precisely to early content degrades. Increasing the context window (C) does not solve attention dilution and may worsen it. Diagnosing a model defect (A) misattributes a known architectural limitation. Premature summarisation (D) is a plausible contributor but is not the primary cause described in the scenario. The proportionate fix is to restructure the pipeline into per-document passes that preserve explicit attribution at each step, rather than accumulating everything into one growing context.

Domain 5 catches many candidates late in preparation because its failure modes are less intuitive than a tool misrouting error. Reserve at least one dedicated study session specifically for Domain 5 questions.

How is the CCAR-F scored, and what should I take from practice results?

The exam is scored on a scale of 100 to 1000, with a passing score of 720. Anthropic does not publish the raw-to-scaled conversion, so we avoid stating a specific question count as the pass mark. On a linear read the threshold falls around 41 to 42 correct of 60 items, but the actual conversion is not linear and candidates should not plan around that approximation.

Your score report shows pass or fail, your scaled score, and percent-correct by domain. That domain-level breakdown is the most actionable output from any practice session. If you score 52% in Domain 1 but 84% in Domain 4, your remediation path is clear. Treat domain scores as diagnostics, not as a single aggregate verdict.

How should I use practice questions effectively rather than just retaking mocks?

Retaking the same practice exam trains pattern recognition on specific stems rather than the underlying reasoning. After each session, we recommend this three-step process:

1. Sort your wrong answers by domain.
2. For each wrong answer, identify which architectural principle the correct answer applied: root-cause tracing, proportionate fix, deterministic over probabilistic, or context preservation.
3. Return to the relevant concept in our library and work through the guided examples before attempting new questions in that domain.

Stale context, attention dilution, and synthesis attribution loss are less intuitive than a tool misrouting error, which is why Domain 5 (15% of the exam) catches candidates who have studied only the higher-weight domains. Allocate time proportional to weight, but do not neglect the tail.

Candidates who build real agents and MCP servers before sitting the exam report that scenario stems feel familiar because they have encountered the failure modes in practice. Reading documentation is useful; implementing a working tool pipeline is more useful. At $125 per attempt, practical preparation is a worthwhile investment before your first sit.

AI Skill Certs is an independent prep platform and is not affiliated with or endorsed by Anthropic.