- Full Mock Exam #1 (60 Qs · 120 min)
- All 10 diagnostic explanations unlocked
- Official Anthropic doc references
INDEPENDENT CCAR-F BLUEPRINT & PRACTICE
Claude Certified Architect (CCAR-F) Mock Exam
Estimate your 5-domain readiness or take a free 10-question CCAR-F diagnostic before the $125 exam.
CCAR-F ARCHITECT DIAGNOSTIC
Benchmark your architecture skills against the 720 bar
Two scenario questions per official CCAR-F domain with instant 1,000-point scoring.
- No login required
- ~10 minutes
- Instant 5-domain score
Free CCAR-F Diagnostic · No Login Required
10-Question Diagnostic Exam
Submit with Unanswered Questions?
You still have unanswered questions.
CCAR-F Estimated Score
Official 5-Domain Breakdown (720 / 1,000 Pass Bar)
Close Your Weak Domains Before the $125 Exam
Unlock full 60-question timed mocks and all option-by-option explanations. One-time purchase, lifetime access.
- All 3 full mocks (180 unique Qs)
- All diagnostic & mock explanations
- Wrong-answer drill mode
CCAR-F 5-Domain Readiness Estimator
Rate your production confidence (1–5) across Anthropic’s five official domains to estimate your 720/1,000 readiness gap.
Agentic loops (stop_reason), multi-agent orchestration, Task/Agent subagents, SDK hooks, and session resumption.
CLAUDE.md hierarchy & @imports, .claude/rules/ path-specific frontmatter, Plan Mode, and headless CI/CD.
Explicit criteria, boundary few-shot examples, tool-use structured outputs, retry loops, and Message Batches.
Strict JSON input_schema design, structured is_error responses, tool_choice modes, and .mcp.json scoping.
Prompt caching breakpoints, compaction, multi-agent error propagation, escalation, and pass^k reliability.
One-Time Purchase · Lifetime Access · No Subscription
Choose Your CCAR-F Mock Plan
Start free with 10 questions, or unlock full 60-question timed simulations once.
- 10 weighted scenario questions
- Instant 1,000-point score & domain chart
- Q1–Q5 full option rationales (1 per domain)
- Full Mock Exam #1 (60 Qs · 120 min)
- All 10 diagnostic explanations unlocked
- Official Anthropic doc references
- All 3 full mocks (180 unique Qs)
- Complete option-by-option explanations
- Wrong-answer drill mode
- Lifetime access, no subscription
What the Official CCAR-F Exam Tests
The official CCAR-F exam runs 60 items in 120 minutes via Pearson VUE with a 720 / 1,000 passing bar and retake waiting periods of 14, 30, and 90 days. Scenarios test production judgment—keeping workflows deterministic, isolating subagent context, enforcing tenant boundaries in MCP, modularizing CLAUDE.md, and diagnosing pass^k reliability gaps. Ready for full 60-question timed drills? Compare one-time Starter ($4.99) and Pro Pass ($9.99) options on our Pricing & Exam Passes page.
Free CCAR-F Sample Questions · Full Option Rationales
3 Sample Questions & 720/1,000 Scoring Guide
Review three high-weight CCAR-F v1.1 scenario questions (Domains 1–3, 67% of exam weight) with full option-by-option elimination rationales below.
Your team is building an autonomous customer support resolution agent using the Claude Messages API. During staging tests, the agent frequently exits early right after emitting an intermediate message such as 'Let me look up your recent billing history' without executing the underlying lookup, because the loop checks whether the response contains non-empty text content and stops if text is present. Additionally, when Claude requests three parallel lookups in a single turn, the current code appends three separate user messages—one per tool result—causing subsequent turns to degrade into sequential single-tool calls. How should you redesign the agentic loop?
Answer & option-by-option explanation
Correct Answer: Option A. In the Claude Messages API, stop_reason == 'tool_use' is the authoritative signal to execute tools and continue the loop, while 'end_turn' signals normal completion. When Claude emits multiple parallel tool_use blocks in one assistant turn, all matching tool_result blocks must be returned together in the immediately following single user message.
- Option A: Correct. Checking stop_reason rather than assistant text prevents premature termination during reasoning turns, and batching all tool_result blocks into one user message preserves valid message alternation and parallel tool-calling behavior.
- Option B: Incorrect. Parsing natural language markers to terminate loops is brittle and unreliable, an iteration cap is only a safety guardrail rather than a primary stopping condition, and splitting parallel tool results across separate user messages violates the expected tool_use/tool_result turn structure.
- Option C: Incorrect. Forcing tool_choice: 'any' on every turn prevents Claude from ever concluding the conversation with a final text response (creating an infinite tool loop), as Claude would never naturally reach stop_reason == 'end_turn'.
- Option D: Incorrect. Stripping assistant text blocks removes Claude's intermediate reasoning from context, and sending separate API requests per tool result fails because the API requires every tool_use_id from the preceding assistant message to be resolved in the immediate next user message.
Official reference: How to implement tool use ↗
An engineering team works in a TypeScript monorepo where React Testing Library conventions apply to *.test.tsx files scattered across apps/web/, packages/ui/, and libs/checkout/, while backend Prisma migration rules apply only to files under services/**/prisma/**. Currently, placing all conventions into the root CLAUDE.md wastes context tokens on unrelated tasks and lowers instruction adherence, and a newly onboarded developer reports they aren't receiving the team's baseline Git commit conventions that the tech lead put in ~/.claude/CLAUDE.md. How should you restructure the repository configuration?
Answer & option-by-option explanation
Correct Answer: Option B. User-level configuration (~/.claude/CLAUDE.md) is private to an individual machine and not shared via Git; team-wide baseline rules belong in the repository's root CLAUDE.md. For conventions spanning scattered file patterns across directories, .claude/rules/*.md files with YAML paths glob frontmatter load rules conditionally only when matching files are accessed.
- Option A: Incorrect. Manually copying ~/.claude/CLAUDE.md creates configuration drift across teammates, and duplicating CLAUDE.md files into dozens of subdirectories containing test files creates severe maintenance overhead.
- Option B: Correct. Version-controlling shared project instructions in root CLAUDE.md solves the onboarding gap, while path-scoped rules in .claude/rules/ eliminate token waste and cross-directory duplication by loading rules automatically based on glob matches.
- Option C: Incorrect. .claude/settings.local.json is gitignored for personal local settings rather than team sharing, and settings JSON configures permissions, hooks, and environment settings rather than markdown coding conventions.
- Option D: Incorrect. Files in ~/.claude/commands/ are user-scoped rather than committed to the repo, and relying on developers to manually invoke slash commands before editing specific file types leads to inconsistent convention enforcement.
Official reference: Manage Claude's memory and CLAUDE.md files ↗
Your team operates an automated PR review workflow powered by Claude. Developers are starting to ignore its comments because 40% are low-value nitpicks about subjective variable naming, formatting already handled by Prettier, or speculative micro-optimizations. An engineer updated the system prompt to add 'Be conservative and only report high-confidence, severe issues,' which reduced total comments by half—but the agent now stays silent on genuine SQL injection risks while still occasionally flagging minor style preferences. What is the most effective prompt engineering fix?
Answer & option-by-option explanation
Correct Answer: Option C. Vague instructions like 'be conservative' or 'high-confidence only' do not define what is in scope, causing uneven suppression of real bugs alongside noise. Explicit categorical boundaries paired with concrete positive and negative few-shot examples reliably align precision and recall.
- Option A: Incorrect. Temperature 0.0 makes token selection greedy but does not clarify ambiguous scope definitions, and negative all-caps exhortations fail to tell the model which categories constitute a false positive.
- Option B: Incorrect. Uncalibrated self-reported numerical confidence scores suffer from the same underlying ambiguity: a model may be '95% confident' that a subjective variable name violates its internal style preference while feeling '80% confident' about a subtle injection path.
- Option C: Correct. Explicitly enumerating reportable defect categories and excluded noise categories removes scope ambiguity, and few-shot examples calibrate judgment on tricky edge cases.
- Option D: Incorrect. Using a smaller model and truncating max_tokens arbitrarily degrades detection of complex security and logic bugs without addressing the prompt's underspecified review criteria.
Official reference: Prompt engineering best practices ↗
CCAR-F v1.1 Scoring Conversion & Blueprint
720 / 1,000 Passing Bar & 5-Domain Weight Table
CCAR-F scales weighted domain performance to 1,000 points (passing threshold: 720 / 1,000, roughly 43 of 60 items answered correctly across the five weighted domains). Under our diagnostic and mock scoring model, each domain contributes proportionally to its official weight: Score = round(1000 × Σ(domain_accuracy × domain_weight)).
| Domain | Official Weight | Scaled Points | Approx. Items (of 60) | Key Tested Topics |
|---|---|---|---|---|
| D1 Agentic Architecture & Orchestration | 27% | 270 pts | ~16 items | stop_reason loops, subagent isolation, PreToolUse hooks, handoffs |
| D2 Claude Code Configuration & Workflows | 20% | 200 pts | ~12 items | CLAUDE.md hierarchy, .claude/rules/ glob rules, Plan Mode, CI/CD |
| D3 Prompt Engineering & Structured Output | 20% | 200 pts | ~12 items | Explicit criteria, XML grounding, tool-use JSON schemas, Message Batches |
| D4 Tool Design & MCP Integration | 18% | 180 pts | ~11 items | Strict input_schema, structured isError payloads, tool_choice, .mcp.json |
| D5 Context Management & Reliability | 15% | 150 pts | ~9 items | Prompt caching prefixes, /compact, calibrated escalation, pass^k |
| Total Exam Blueprint | 100% | 1,000 pts | 60 items · 120 min | 720 / 1,000 passing score |
Architect Exam FAQ
What is included free versus Starter and Pro Pass?
The 10-question diagnostic (Q1–Q5 full rationales, 1 per domain) and 5-domain estimator are free with no login. Starter ($4.99 one-time) unlocks Mock #1 and all 10 diagnostic explanations; Pro Pass ($9.99 one-time) unlocks all 3 mocks (180 Qs) and wrong-answer drills.
What are the official CCAR-F domain weights and passing score?
CCAR-F weights D1 Agentic Architecture (27%), D2 Claude Code Workflows (20%), D3 Prompting & Structured Output (20%), D4 Tool & MCP (18%), and D5 Context & Reliability (15%), with a 720/1,000 passing bar.
When does the exam countdown timer start?
The timer never starts on page load. It begins only after you choose Timed Exam Mode and click Start Test Now.
Is AI Cert Mock affiliated with Anthropic?
No. All diagnostics, scenario questions, and distractor rationales are original unofficial study materials created independently of Anthropic.