Mock exam 1: Developer – Foundations
A timed practice exam in the official style. Unlike the practice questions, the domains are hidden and the answers sit at the end, so this measures readiness rather than teaching. Original questions written against the public blueprint; not items from the live exam, which is covered by a non-disclosure agreement.
How to take it. 15 questions, 30 minutes, closed book: no notes, no documentation, no Claude. Answer every question, since unanswered items score zero. Question weighting follows the blueprint, so Applications and Integration dominates. Score yourself with the key and the readiness table afterwards.
Questions
1. Your service makes a Messages API call and the response carries stop_reason "end_turn" with a tool_use block absent. What does your loop do?
- B. Execute the last requested tool again
- A. Treat the turn as complete and return the content to the caller
- C. Re-send the request with higher max_tokens
- D. Append an empty tool result and continue the loop
2. A nightly job enriches 200,000 product descriptions. Results are consumed the next morning. Which design minimizes cost?
- A. Synchronous calls fanned out across many workers
- B. The Message Batches API with results collected within its processing window
- C. Synchronous calls with a smaller model and reduced max_tokens
- D. One request containing all descriptions
3. Which change makes a repeated 10,000-token system prompt cheaper and faster without losing content?
- A. Moving the prompt to the end of the request
- C. Placing the stable prompt first and enabling prompt caching so the prefix is reused
- B. Compressing the prompt text with abbreviations
- D. Sending the prompt only on the first request of a session
4. Your extraction endpoint occasionally receives JSON with a missing required field, and downstream code throws. What is the sturdiest production pattern?
- A. Wrap the parse in a try block and return an empty object on failure
- D. Enforce a schema through tool use, validate the parsed result, and retry once with the validation error appended
- B. Instruct the model more firmly in the system prompt
- C. Post-process with a regular expression that inserts missing fields
5. A teammate proposes an autonomous agent for a pipeline whose five steps are identical for every input. What is the strongest technical objection?
- B. Agents are unsupported in production
- A. A known, fixed path does not need model-chosen control flow; deterministic orchestration is cheaper, testable, and auditable
- C. Agents cannot call tools reliably
- D. The pipeline would need a larger context window
6. Which is the correct use of a subagent in an agent system?
- A. To run the same prompt twice for consensus
- B. To isolate a self-contained subtask so its verbose intermediate context does not fill the parent's window
- C. To bypass rate limits
- D. To provide a fallback if the main model is unavailable
7. An agent that summarizes user-submitted URLs must never be able to trigger a refund. Where is that guaranteed?
- A. In the system prompt, stated clearly
- C. By not granting the agent the refund tool, and enforcing that with tool configuration or hooks
- B. By instructing the model to confirm before refunds
- D. By using a model known for instruction following
8. Your team pins model versions in production. What is the primary reason?
- A. Pinned models are cheaper
- D. Model updates can change behavior, so upgrades become deliberate, evaluated changes rather than surprises
- B. Pinning increases the context window
- C. Pinning is required by the API
9. Which describes the correct relationship between an MCP server and the applications that use it?
- B. Each application embeds its own copy of the server code
- A. The server exposes tools and resources that multiple clients connect to, maintained independently of any one application
- C. The server replaces the Messages API
- D. The server must run in the same process as the model
10. A long-running assistant begins contradicting itself and forgetting recent constraints. Inspection shows the context is nearly full of verbose tool outputs. What is the correct remedy?
- A. Increase temperature to encourage variety
- B. Prune and summarize tool outputs, compact history, and isolate subtasks so the working context stays lean
- C. Repeat the system prompt after every turn
- D. Reduce max_tokens on each response
11. When is few-shot prompting the right tool?
- A. When the model needs more factual knowledge
- C. When output format or edge-case handling is inconsistent and examples can demonstrate the target behavior
- B. When latency must be reduced
- D. When the context window is too small
12. A page retrieved by your agent contains hidden text instructing it to reveal its system prompt. Which mitigation actually works?
- A. Asking the model in the system prompt to ignore such instructions
- D. Treating retrieved content as untrusted data, separating it from instructions, and gating sensitive capabilities behind guardrails
- B. Raising temperature so behavior is less predictable
- C. Filtering pages that contain the word "ignore"
13. Which statement about the Message Batches API is correct?
- B. It supports multi-turn tool calling within a batch
- A. It processes large asynchronous workloads at reduced cost within a processing window, correlating requests by custom_id
- C. It guarantees sub-second latency
- D. It requires a dedicated endpoint per request
14. Your CI pipeline runs Claude Code non-interactively and must parse its output. Which invocation fits?
- A. Interactive mode with output piped to a file
- B. The print flag with a JSON output format and a schema, so results are machine-readable and validated
- C. Plan mode with a human approver
- D. A wrapper that screen-scrapes the terminal
15. Traces show your assistant reasons correctly over the documents it receives, but frequently receives documents unrelated to the question. Where is the defect?
- A. The model, which should be upgraded
- C. The retrieval layer that selects the documents
- B. The system prompt, which needs stronger accuracy language
- D. The output parser
Answer key
| # | Answer | Domain | Why |
|---|---|---|---|
| 1 | A | Agents and Workflows | "end_turn" means the model has finished; "tool_use" is the signal to execute a tool and continue the loop |
| 2 | B | Applications and Integration | Latency-tolerant bulk work is the Batches API's purpose: roughly half the cost within a 24-hour window |
| 3 | C | Applications and Integration | Caching reuses a stable prefix, cutting time-to-first-token and per-request cost. Position matters: stable content first |
| 4 | D | Prompt and Context Engineering | Schema enforcement plus a validation-retry loop treats model output as untrusted input and lets the model self-correct |
| 5 | A | Agents and Workflows | Autonomy is justified when the path is unknown. A fixed path belongs in deterministic code that can be tested |
| 6 | B | Agents and Workflows | Subagents delegate and isolate context; that isolation is their main architectural value |
| 7 | C | Security and Safety | Capability the agent does not hold cannot be misused. Prompts and confirmations are not enforcement |
| 8 | D | Model Selection and Optimization | Pinning makes model behavior a controlled dependency so upgrades pass through evaluation |
| 9 | A | Tools and MCPs | An MCP server is a reusable, independently maintained capability that multiple clients share |
| 10 | B | Prompt and Context Engineering | Context drift and bloat are managed by pruning, compaction, and isolation |
| 11 | C | Prompt and Context Engineering | Few-shot examples fix format and edge-case behavior far better than more instructions |
| 12 | D | Security and Safety | Injection defense is structural: separate untrusted content and apply least privilege to consequential tools |
| 13 | A | Applications and Integration | Reduced cost, asynchronous window, custom_id correlation, and no multi-turn tool calling inside a batch |
| 14 | B | Claude Code | Headless CI use means non-interactive invocation with structured, schema-validated output |
| 15 | C | Eval, Testing, and Debugging | Traces localize the fault to the integration layer: correct reasoning over wrong inputs is a retrieval defect |
Interpreting your score
| Raw score | Reading |
|---|---|
| 13 to 15 | Comfortable. Review misses and book the exam |
| 10 to 12 | Close. Target the domains where you lost points, especially anything in Applications and Integration, which is a third of the real exam |
| 7 to 9 | More study needed. Build something that exercises the API, a tool, and structured output, then return |
| 6 or fewer | Start from the blueprint and the official exam guide, then the prep path |
This scale is a study aid, not a predictor. The real exam reports a scaled score from 100 to 1,000 with 720 to pass.
Tally misses by domain; two or more in one domain marks your next study target, as the real score report's percent-correct breakdown would.
Written by the maintainer for self-assessment. Study guide · Study notes · Mock 2 · Mock 3 · Practice questions · Repository index