Test your exam readiness

Sample Claude Certified Developer practice questions

6 free sample questions from the practice bank, with the correct answer and the reasoning.

1. Question 1 · Domain 1 — Agents and Workflows

   Support Ticket Triage Agent

   A support-triage loop decides when to stop by scanning the assistant's text for the phrase "issue resolved." In production, the loop sometimes stops while the customer is mid-conversation and sometimes runs long after the ticket is actually closed. What should replace the text scan?

   • A.A shorter trigger phrase such as "done", since long phrases are less reliable for Claude to reproduce consistently
   • B.A regex that fuzzy-matches several synonyms of "resolved" to widen the net of accepted completion phrases
   • C.Checking stop_reason === "end_turn" on the API response, since that is the deterministic signal that the model has finished its turn Correct answer
   • D.Lowering temperature to 0 so Claude's wording becomes consistent enough for reliable text matching

   Why: The API's stop_reason field is a structural contract, not a phrase the model chooses to say — end_turn means the model has finished normally, independent of exact wording. Text inspection fails because natural-language phrasing varies run to run. Why the other options are wrong: a shorter or fuzzier phrase is still text inspection and still unreliable. Lower temperature reduces wording variance but does not guarantee the phrase appears at all — it does not turn prose into a deterministic signal.

2. Question 2 · Domain 2 — Applications and Integration

   Enterprise Deployment Decision

   A team wants to share a specific agent behavior (a support bot's system prompt, tool definitions, example flows, and README) so other teams can adopt it without starting from scratch. Which accelerator format fits, and why?

   • A.An eval suite, because validation is always the primary deliverable of an accelerator
   • B.An agent template, because the pattern being shared is a specific agent behavior — the format built for exactly this purpose Correct answer
   • C.An MCP server package, because any shareable Claude artifact must be exposed via MCP
   • D.A deployment platform, because sharing requires choosing Bedrock or Vertex AI first

   Why: The three standard accelerator formats map to what's being shared: an agent template is for a specific agent behavior (system prompt + tools + example flows + README) — exactly the case described. Why the other options are wrong: an eval suite packages test cases and grading logic, not a behavior pattern itself. An MCP server package is for reusable tools, not a full agent behavior. Deployment platform choice is unrelated to which accelerator format best packages a shareable pattern.

3. Question 3 · Domain 3 — Claude Code

   Onboarding a New Engineering Team

   A new hire is running unfamiliar, exploratory Claude Code sessions on a production repo. Following the principle of least privilege, which permission mode should they start in?

   • A.bypassPermissions, so they can move quickly while learning the codebase
   • B.auto, so file writes and most commands proceed without repeated prompts
   • C.default (or plan for design-only exploration), which reads files and asks before writing or running commands — the minimum privilege needed for unfamiliar, exploratory work Correct answer
   • D.dontAsk, since it's less permissive than auto

   Why: Least privilege means giving Claude the minimum permissions needed for the current task — for unfamiliar, exploratory work, default (or plan, which doesn't even allow writes) is the appropriate starting point; broader modes are earned once the task and trust level are established. Why the other options are wrong: bypassPermissions and auto grant far more autonomy than exploratory work requires. dontAsk is actually more permissive than auto (everything except bypassing safety), not less — ranking it as more restrictive is incorrect.

4. Question 4 · Domain 4 — Eval, Testing, and Debugging

   Invoice Extraction Service

   A team is about to start building an invoice-extraction feature. One engineer wants to write the system prompt first and "see how it does," then write eval cases afterward based on what looks reasonable. What is wrong with this order?

   • A.Nothing — eval cases are more accurate when written against real model outputs
   • B.Eval cases must always be written by a separate team from the one writing the prompt
   • C.Writing the prompt first means success criteria get defined around whatever the prompt already produces, optimizing for outputs that look good rather than outputs that are actually correct Correct answer
   • D.Eval cases can only be authored after a model version is pinned, so prompt-first is the only viable order

   Why: The eval precedes the prompt. Defining success criteria before writing the system prompt keeps the target independent of whatever the first draft happens to produce. Writing the prompt first anchors "correct" to "what this prompt already does," which quietly optimizes for familiar-looking output instead of the actual task requirement. Why the other options are wrong: eval cases don't need to be derived from model outputs to be accurate — they should be derived from the actual task requirement. There's no rule requiring separate teams. Model pinning is a deployment concern, unrelated to which comes first, the prompt or the eval.

5. Question 5 · Domain 5 — Model Selection and Optimization

   High-Volume Ticket Classifier

   A ticket-classification prompt processes mostly English text but occasionally includes Japanese customer quotes. Why does the Japanese content disproportionately affect the token budget?

   • A.Japanese text is rejected by the tokenizer and must be translated to English first
   • B.Non-Latin scripts like Japanese are split into more tokens per character (roughly 2–3 tokens per character) than English, which averages about 1.3 tokens per word Correct answer
   • C.Japanese text always triggers extended thinking mode, which consumes additional tokens
   • D.The tokenizer charges double for any request containing mixed scripts

   Why: Claude's BPE tokenizer represents common English words as roughly one token each (about 1.3 tokens per word), but non-Latin scripts like Chinese, Japanese, or Korean are split much more finely — roughly 2–3 tokens per character — making the same amount of visible text cost meaningfully more of the token budget. Why the other options are wrong: the tokenizer doesn't reject non-Latin scripts or require translation. Extended thinking is a separate, explicitly-enabled mode unrelated to input script. There's no mixed-script surcharge — token cost follows from how the BPE tokenizer splits the actual characters.

6. Question 6 · Domain 6 — Prompt and Context Engineering

   Compliance Review Assistant

   A compliance-review assistant's output is consistently off-topic and speaks in an overly casual tone, ignoring the intended reviewer persona. Which of the four prompting techniques is missing?

   • A.XML tags — instructions and content are getting confused
   • B.Few-shot examples — the model doesn't know the target format
   • C.System prompt — role, persona, and constraints haven't been set at the operator level Correct answer
   • D.Output constraints — the response isn't structured as JSON

   Why: Off-topic, wrong-persona output is the specific symptom the diagnosis table maps to a missing or weak system prompt — the operator-level context that establishes role, tone, and constraints for the entire conversation. Why the other options are wrong: XML tags solve mixed-content confusion (instructions vs. data), not tone or persona drift. Few-shot addresses unclear format, not identity/persona. Output constraints govern structure (JSON vs. prose), unrelated to the casual-tone symptom described.

What this practice exam covers

415 questions · 8 exam domains · 1 mock exam

• Domain 1 — Agents and Workflows
• Domain 2 — Applications and Integration
• Domain 3 — Claude Code
• Domain 4 — Eval, Testing, and Debugging
• Domain 5 — Model Selection and Optimization
• Domain 6 — Prompt and Context Engineering
• Domain 7 — Security and Safety
• Domain 8 — Tools and MCPs

More practice for this exam

Claude Certified Developer: the 53-question quiz

Questions about this practice exam

Are these the official Claude Certified Developer exam questions?

No. They are practice questions written against the published exam domains. This site is not affiliated with or endorsed by Anthropic.

How many practice questions are there?

415 questions across 8 exam domains including 1 full mock exam.

Do I need an account to practise?

No. You can take the practice questions without an account and your scores save in your browser. A free account adds the full answer review, gap analysis and a completion certificate.

Do the questions explain the answers?

Yes. Every question comes with an explanation of the correct answer.