The CCAR-F exam tests whether you can build reliable systems with Claude, not whether you can recite feature lists. Every question sits inside a production scenario and asks what you would do next. This guide covers the format, the five domains, the scenario draw, and a study plan that matches how the exam actually asks questions.
Everything here comes from the published exam guide (v1.0, July 2026). Check the guide itself before you book, because details can change between versions.
The format at a glance
- Items: 60.
- Time: 120 minutes, which works out to 2 minutes per item.
- Item types: multiple-choice and multiple-response. Each multiple-response item tells you how many answers to select.
- Scoring: scaled, with a pass mark of 720 on a 1,000-point scale.
- Delivery: proctored, delivered through Pearson VUE (online or at a test center), per the published guide.
Two minutes per item sounds generous. It isn't, once you account for reading a scenario setup and comparing four plausible options. Every item is framed by a scenario, but some read faster than others, so move quickly on those and slow down on the hard judgment calls.
The scenario structure: 4 of 6
The exam is scenario-based. Each sitting draws 4 scenarios at random from a bank of 6, and each scenario frames a set of questions. The six scenarios are:
- Customer support resolution agent. An Agent SDK agent handling returns, billing disputes and account issues through MCP tools such as
get_customer,lookup_order,process_refundandescalate_to_human. - Code generation with Claude Code. Custom slash commands, CLAUDE.md setup, and plan mode versus direct execution.
- Multi-agent research system. A coordinator delegates to search, analysis, synthesis and reporting subagents to produce a cited report.
- Developer productivity. An agent that explores unfamiliar codebases with built-in tools (Read, Write, Edit, Bash, Grep, Glob) plus MCP servers.
- Claude Code in CI/CD. Automated review, test generation and PR comments with minimal false positives.
- Structured data extraction. Pulling data from unstructured documents, validating it against JSON schemas, and feeding downstream systems.
The draw matters for planning. If the draw is uniform, any single scenario has a 2-in-3 chance of appearing. If you skip two scenarios, the odds that at least one of them shows up are 14 in 15. You cannot study around the draw, so prepare all six.
The five domains and what they test
- Agentic Architecture & Orchestration: 27%
- Tool Design & MCP Integration: 18%
- Claude Code Configuration & Workflows: 20%
- Prompt Engineering & Structured Output: 20%
- Context Management & Reliability: 15%
Domain 1: Agentic Architecture & Orchestration (27%)
The heaviest domain. It covers the agentic loop, coordinator–subagent design, hooks and session handling.
In practice it asks whether you know what controls the loop: continue on stop_reason: "tool_use", stop on "end_turn". It also tests that subagents do not inherit the coordinator's history, so context must be passed explicitly. Expect questions on when a prompt instruction is not enough and a hook or prerequisite gate must enforce the rule.
Domain 2: Tool Design & MCP Integration (18%)
This domain is about making tools the model can choose correctly. Tool descriptions are the main signal for tool selection, so vague or overlapping descriptions cause misrouting. It also covers structured MCP errors, scoping each agent to a small set of role-relevant tools, tool_choice settings, and where MCP configuration lives.
Domain 3: Claude Code Configuration & Workflows (20%)
This is the practical setup domain: the CLAUDE.md hierarchy, project versus personal commands and skills, and path-specific rules in .claude/rules/. It also covers plan mode versus direct execution and running Claude Code non-interactively in CI with -p. A typical question describes a teammate who isn't getting instructions and asks why.
Domain 4: Prompt Engineering & Structured Output (20%)
Explicit criteria, few-shot examples, JSON-schema output through tool use, validation and retry, and batch processing. The key distinction is that schema-enforced output removes syntax errors but not semantic ones. It also tests when the Message Batches API fits (overnight reports) and when it doesn't (blocking pre-merge checks).
Domain 5: Context Management & Reliability (15%)
Keeping facts intact over long sessions, escalation triggers, error propagation across agents, human review, and source provenance. In practice it asks whether you'd keep amounts and dates in a persistent "case facts" block, and whether you can tell a failed search from a search that found nothing.
How to study
Learn the anti-patterns by name
The exam guide names many anti-patterns directly. Learn them as a list, because wrong options tend to look like them:
- Detecting loop completion by parsing the model's text, or using an iteration cap as the main stop mechanism.
- Relying on a prompt instruction where compliance is critical, such as identity checks before a refund.
- Giving one agent too many tools, or giving a synthesis agent tools outside its role.
- Returning a generic "operation failed" error, or returning an empty result as a success.
- Using sentiment or self-reported confidence as an escalation trigger.
- Asking the session that wrote the code to review it, instead of an independent instance.
- Expecting "be conservative" to improve precision.
When you review a missed question, name which anti-pattern the distractor used. That habit transfers to questions you haven't seen.
Weight your time by domain
Domains 1, 3 and 4 make up 67% of the exam. If weights map to item counts, Domain 1 alone is about 16 of the 60 items. Spend study time in rough proportion, but don't skip Domain 5: at 15%, it is still about 9 items.
Use timed mocks to train pacing
Take full-length mocks under real conditions: 60 items, 120 minutes, no notes. Untimed practice hides the real problem, which is spending four minutes on one judgment call and rushing the last ten items. After each mock, record your score by domain, rather than only the total.
Drill weak domains, then retest
A total score tells you whether you're ready. A domain breakdown tells you what to fix. Take your weakest domain, re-study the matching section of the exam guide, then re-drill only the questions you missed. Retest with a fresh full mock rather than repeating the same one, since a repeated mock measures memory, not understanding.
Common traps
- The bigger-model answer. If an option fixes a design flaw by upgrading the model or adding context, it is usually a distractor.
- The prompt-only fix for a hard rule. If the rule must hold every time, look for a hook or gate.
- Out-of-scope options. The guide lists topics it does not test, including fine-tuning, embeddings, vision and computer use. An answer that leans on them is a warning sign.
- Multiple-response miscounts. Read how many answers the item asks for before you start eliminating.
Practice with Ready Room
Ready Room for CCAR-F gives you full mock exams with a missed-question re-drill and a domain heat map, so you can see which domains need work. Sigma Foundry makes this app. It's an unofficial practice resource, not affiliated with Anthropic.
Prefer Udemy? The same six-exam CCAR-F set is on Udemy as Claude Architect Foundations (CCAR-F) Practice Exams: six timed 60-question mocks, every option explained. It's $12.99 at launch with that link (list $19.99) through December 5.
FAQ
How many questions are on the CCAR-F exam?
60 items, multiple-choice and multiple-response, in 120 minutes.
What score do I need to pass?
A scaled score of 720 out of 1,000, per the published exam guide (v1.0).
Will I see all six scenarios?
No. Each sitting draws 4 of the 6 at random, so prepare for all six.
Which domain matters most?
Agentic Architecture & Orchestration, at 27%.