Claude Certification Passing Score: CCAR-F Complete Guide

What is the claude certification passing score?

The claude certification passing score for the Claude Certified Architect, Foundations exam (CCAR-F) is 720 on a scaled score of 100 to 1,000. Anthropic does not publish the raw-to-scaled conversion formula, so we will not state an exact question count as the pass mark. What we can say is that on a rough linear read, 720 out of 1,000 implies somewhere in the range of 41 to 42 correct answers out of 60, but the actual conversion may differ. Plan to answer every question well rather than counting towards a minimum.

Your score report gives you three things: a pass or fail verdict, your scaled score, and percent-correct broken down by domain. That domain breakdown is the most actionable part of the report for anyone who needs to retake.

How is the CCAR-F exam structured?

The exam is 60 items delivered in a 120-minute window, either online-proctored or at a Pearson VUE test centre. Every item is scenario-based and tests practical judgement, not recall. Items are multiple-choice or multiple-response; when a question requires more than one answer, it states exactly how many to select.

One structural feature that surprises many candidates: each sitting draws 4 scenarios at random from a bank of 6. That means you cannot predict which four you will face. Preparation must be broad enough to handle any combination.

Which domains carry the most weight?

The five domains and their exam weights are published in the official exam guide. Domain 1 alone accounts for more than a quarter of the exam, making Agentic Architecture & Orchestration the single highest-leverage area to study.

Domains 3 and 4 are tied at 20% each, making them jointly the second-largest block. Together, Domains 1, 3, and 4 account for 67% of the exam. If you are short on study time, that is where to concentrate first.

Domain 5 at 15% is the smallest slice, but do not neglect it: context management failures are a common root cause of the reliability anti-patterns the exam tests explicitly.

Where do the biggest scoring gains come from?

Answer-first: the biggest gains come from mastering the scenario logic in Domain 1 and the configuration mechanics in Domain 3, because those two domains together represent 47% of the exam and both reward deterministic, root-cause thinking over probabilistic guessing.

Domain 1: Agentic Architecture & Orchestration (27%)

The exam tests whether you can design and debug multi-agent systems under realistic constraints. Key concepts include hub-and-spoke architecture, subagent context isolation, and agentic loop anti-patterns. Scenarios frequently involve a coordinator that must route work to specialised subagents, handle partial failures, and decide when to escalate to a human. The exam consistently rewards deterministic solutions over probabilistic ones when stakes are high.

Domain 3: Claude Code Configuration & Workflows (20%)

This domain goes deep on CLAUDE.md files, .mcp.json configuration, the three-level configuration hierarchy, and version control implications. Expect questions about which configuration scope applies in a given directory context, and what happens when project-level and user-level settings conflict.

Domain 4: Prompt Engineering & Structured Output (20%)

Scenarios here test whether you can diagnose why a prompt produces inconsistent output and choose the proportionate fix. The exam rewards candidates who can distinguish between a problem that needs a schema change, a few-shot example, or a system prompt revision, rather than reaching for the most complex solution available.

Domain 2: Tool Design & MCP Integration (18%)

Questions test tool description quality, error propagation patterns, and the isError flag. A common scenario presents a tool that is being misrouted and asks you to identify the root cause. Knowing tool descriptions as a selection mechanism is essential here.

Domain 5: Context Management & Reliability (15%)

This domain covers stale context, attention dilution, and session management strategy. Scenarios often involve an agent that degrades over a long session and ask you to choose between resuming, forking, or starting fresh.

What anti-patterns does the exam actually test?

The exam is explicit about testing anti-patterns, not just correct patterns. Here are the categories that appear most frequently across domains:

Agentic loop failures. Questions present a loop that terminates prematurely or runs indefinitely and ask you to identify the cause. The stop_reason field is a common diagnostic tool in these scenarios.

Narrow decomposition. A coordinator that decomposes a task too narrowly misses cross-cutting concerns. The exam tests whether you recognise narrow decomposition failure and can propose a fix that addresses the root cause rather than patching symptoms.

Tool overload. Giving an agent too many tools degrades selection accuracy. The exam tests whether you know when to split tools, scope them, or use tool_choice configuration to constrain selection.

Silent suppression. An agent that swallows errors rather than propagating them is a reliability anti-pattern. Expect scenarios where you must distinguish between a valid empty result and an access failure.

Stale context. Long sessions accumulate irrelevant context that dilutes attention. The exam tests whether you know when to inject a summary versus forking to a fresh session.

How should you sequence your study?

We recommend a domain-weight-proportionate approach. Spend roughly the same fraction of your study time on each domain as that domain contributes to the exam score.

1. Start with Domain 1 (27%). Build a mental model of multi-agent orchestration before touching anything else. Work through coordinator responsibilities, dynamic subagent selection, and error routing.
2. Move to Domains 3 and 4 together (20% each). Configuration and prompt engineering share a common theme: precision over improvisation. Study them in parallel because Claude Code scenarios often involve both.
3. Cover Domain 2 (18%). Tool design and MCP integration build naturally on the agent architecture you studied in Domain 1.
4. Finish with Domain 5 (15%). Context management is the smallest domain but underpins reliability across all others. Study it last so the concepts land in a richer mental model.

Because each sitting draws 4 of 6 scenarios at random, you cannot safely skip any scenario type. Treat all six as in-scope.

What does the scenario bank mean for preparation?

Each of the 6 scenarios in the bank maps to a realistic production context: multi-agent research pipelines, support escalation workflows, CI/CD automation, data extraction and validation, code review agents, and configuration management. You will face 4 of these at random.

The practical implication is that you need broad coverage, not deep specialisation in two or three scenario types. A candidate who has studied only the research pipeline and CI/CD scenarios is exposed if the exam draws support escalation and data extraction instead.

Our concept library at /concepts maps 174 atomic concepts to all five domains and 30 task statements, which gives you a structured way to verify you have covered every scenario type without gaps.

How does the scoring scale work in practice?

The 100-to-1,000 scale is not linear in the way a percentage score is. Anthropic applies a raw-to-scaled conversion that we do not have visibility into. What this means practically:

• A score of 719 is a fail regardless of how close it feels to 720.
• A score of 720 is a pass regardless of which questions you got wrong.
• The domain percent-correct breakdown on your report tells you where you lost points, which is the data you need for a retake strategy.

Do not target "just enough to pass." The scenario-based format means a single misread scenario can cost you several items at once. Build a comfortable margin.

Is there a practice exam that mirrors the real format?

AI Skill Certs is an independent prep platform (not affiliated with or endorsed by Anthropic). Our CCAR-F practice exams are scored on the same 100-to-1,000 scale with 720 as the passing bar, and they use the same 60-question, scenario-based format as the real exam. The adaptive engine uses Bayesian Knowledge Tracing with a 0.90 mastery threshold to identify gaps before you sit the real exam.

If your practice score is consistently above 780, you have a reasonable buffer. If you are scoring between 700 and 750, the domain breakdown will tell you exactly which domain to revisit.