Skip to main content

Practice exam · Anthropic

Claude Certified Architect, Foundations CCAR-F practice exam

65
questions
120
minutes
72%
to pass the practice exam
70
questions in the bank

Short answer

The Claude Certified Architect, Foundations CCAR-F practice exam has 65 questions in 120 minutes, with a 72% passing mark, as configured here. The fastest way to know whether you are ready is to sit a full exam on the clock and read your score by domain, not only the total.

Domains covered

  • Agentic Architecture & Orchestration
  • Claude Code Configuration & Workflows
  • Prompt Engineering & Context Management
  • Tool Design & MCP
  • Evaluation & Cost Planning
  • Safety, Governance & Deployment

Sample questions, with explanations

Three questions from the bank, open to read.

Question 1 · Evaluation & Cost Planning

A document-extraction pipeline processes a million documents monthly. A small model handles most documents correctly at low cost but fails on complex ones; the top-tier model handles nearly everything but is roughly an order of magnitude more expensive. Extraction outputs can be checked cheaply (schema validation, arithmetic consistency, confidence signals). What architecture captures most of the quality at a fraction of the cost?

  1. Average the small and large models' outputs on every document to blend cost and quality
  2. Send documents to a random model weighted 90/10 toward the small one to cap spend probabilistically
  3. Run the top-tier model on everything, since extraction errors are unacceptable at any rate
  4. An escalation cascade: run the small model first, apply the cheap automated checks to its output, and escalate only failing or low-confidence documents to the top-tier model — so the expensive model runs on the small fraction that actually needs it ✓ (correct)
Why this is correct

Cheap, reliable output verification is the key enabler: it converts model choice from a static routing decision into a per-document escalation policy, so the expensive model's cost is incurred only on the minority of documents where the small model demonstrably struggled — typically retaining near-top-tier quality at a small multiple of small-model cost. Averaging structured extractions is ill-defined and runs both models on every document, maximizing cost. Random weighted routing ignores document difficulty entirely, sending hard documents to the weak model 90% of the time. Top-tier-everywhere ignores that the cascade reaches equivalent final quality — failures are caught and escalated — while spending an order of magnitude more.

Question 2 · Claude Code Configuration & Workflows

Every release, an engineer types the same multi-paragraph prompt into Claude Code: gather merged PRs since the last tag, categorize them, and draft release notes in the team's format. The team wants this workflow reusable, parameterized by version, and shared with everyone via the repository. What is the idiomatic Claude Code mechanism?

  1. Save the prompt in a personal notes app and paste it each release
  2. Ask Claude at session start to memorize the procedure for all future sessions
  3. Create a custom slash command (or skill) checked into the repository — e.g., a .claude/commands/release-notes.md file with argument support — so any teammate can invoke the whole workflow with one command ✓ (correct)
  4. Hardcode the full procedure into CLAUDE.md so it loads into every session
Why this is correct

Custom slash commands (and skills) exist for exactly this: a prompt template checked into the repo under .claude/commands/ becomes a shared, versioned, argument-accepting workflow that any teammate invokes by name. A personal notes app keeps the workflow siloed with one engineer and invites drift. Asking Claude to 'memorize' a procedure does not persist across sessions by itself — durable instructions must live in files. Embedding the whole procedure in CLAUDE.md taxes every session's context with content needed only at release time, which is the wrong loading model for an occasional workflow.

Question 3 · Safety, Governance & Deployment

A data-analysis product lets an agent write and execute Python against user-uploaded datasets. The runtime is currently the API server's own environment. Security review flags that generated code could read environment secrets, reach internal services, or exfiltrate data. What is the required execution architecture?

  1. Static-analysis scanning of generated code for dangerous calls before executing it in place
  2. A system-prompt instruction forbidding the model from generating code that accesses the network or environment
  3. Human review of every generated script before execution, keeping the current runtime
  4. Execute generated code only inside an isolated sandbox — an ephemeral, least-privilege container with no credentials or secrets, filesystem access limited to the session's data, and default-deny egress with only required destinations allowed ✓ (correct)
Why this is correct

Model-generated code must be treated as untrusted code, and the control that actually bounds it is the execution environment: an ephemeral sandbox holding no secrets, seeing only the session's own data, and unable to reach the network except through an explicit allowlist makes exfiltration and lateral movement structurally impossible rather than merely discouraged. Static analysis of arbitrary Python is reliably evadable (dynamic imports, encodings, indirect calls) and the code still runs beside the secrets when it slips through. Prompt instructions do not bind a model influenced by injected data content. Human review of every script destroys the product's interactivity and still executes approved code in the dangerous environment.

Take the full practice exam

  • 65 questions drawn from the bank, with a 120-minute timer
  • Score by domain and a review of every question
  • An AI tutor to talk through each question you miss
  • What you miss feeds your study plan
Start the practice examStudy the path first

Independent, unofficial practice exam. Anthropic and the exam name are trademarks of their owners; UpSkillScore is not affiliated with them. Format and passing mark are the exam's configuration in UpSkillScore, not an official statement.

Frequently asked questions

How many questions does the Claude Certified Architect, Foundations CCAR-F practice exam have?

It has 65 questions and a 120-minute limit. The passing mark of the practice exam is 72%.

Is this the official Claude Certified Architect, Foundations CCAR-F exam?

No. It is an independent practice exam built to match the structure of the real one. It does not replace the official exam.

Do I need an account to take it?

You can read the structure and sample questions without an account. Taking the full timed exam, scoring by domain and the AI tutor need one.