Mock exam 1: Architect – Professional
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. Several questions test judgment about stakeholders and governance, which carry real weight on this exam. Score yourself with the key and the readiness table afterwards.
Questions
1. A retailer wants Claude to answer product questions from a 40,000-document catalog that changes daily. Which architecture fits?
- A. Fine-tune a model on the catalog nightly
- C. Retrieval over an index refreshed as the catalog changes, with the model answering from retrieved context
- B. Load the catalog into the system prompt
- D. An autonomous agent that browses the public website
2. After a re-index, your retrieval system returns semantically similar but commercially wrong products. Which investigation comes first?
- A. The generation prompt
- D. Chunking and embedding configuration, and whether the index was rebuilt consistently
- B. The model version
- C. The user interface
3. An agent used by 300 employees exposes 45 tools across eight systems. Selection quality is degrading. What is the architectural response?
- B. Shorten all tool descriptions
- A. Scope toolsets per task or role, with progressive discovery for the rest
- C. Merge tools into a few multi-mode tools
- D. Replace the agent with a larger model
4. Which evaluation design best fits a summarization feature where quality is subjective?
- A. Exact match against reference summaries
- B. A rubric covering the failure modes that matter, applied by model graders and periodically audited by humans on a labeled set
- C. Average output length compared to references
- D. User thumbs-up counts alone
5. Before a prompt change reaches all users, what is the disciplined sequence?
- A. Ship it and monitor complaints
- C. Regression evaluation, then a controlled A/B against the current prompt on predefined metrics, then rollout
- B. Team vote followed by rollout
- D. Rollout with a feedback form
6. A hospital system requires that protected health information never reach the model provider, while clinicians need free-text search. Which design satisfies both?
- A. A policy instructing clinicians not to type identifiers
- D. A de-identification layer that removes or tokenizes PHI before the call, with local re-association and auditing
- B. A system prompt instructing the model to disregard PHI
- C. Restricting the feature to structured queries only
7. Which is the correct application of human-in-the-loop review at scale?
- B. Review every output regardless of risk
- A. Route by confidence and consequence, so low-confidence and high-impact cases reach humans while routine cases flow through
- C. Review a fixed 10% sample chosen at random
- D. Review only what users complain about
8. An executive sponsor demands a single accuracy number for a system whose performance varies sharply by document type. What does a competent architect present?
- A. The highest observed number
- B. Performance broken out by document type with the aggregate, the measurement method, and what the business outcome is after human review
- C. A refusal to quantify
- D. An average weighted to look favorable
9. Your design must satisfy GDPR for EU customer data. Which requirement most directly shapes the architecture?
- A. Response latency targets
- C. Data minimization, purpose limitation, and the ability to delete personal data on request
- B. The choice of programming language
- D. The number of model providers
10. Which is the correct reason to prefer a workflow over an agent for a document intake process?
- A. Workflows are always more accurate
- D. The processing path is known in advance, so deterministic orchestration is cheaper, testable, and auditable
- B. Agents cannot process documents
- C. Workflows use less context
11. An application sends an identical 12,000-token policy preamble on every request. Which optimization addresses both latency and cost?
- B. Splitting the preamble across two requests
- A. Ordering the static preamble first and enabling prompt caching
- C. Summarizing the preamble with the model at request time
- D. Moving the preamble into a few-shot block
12. What belongs in an architecture decision record for a Claude system?
- A. The final design only
- B. The decision, the alternatives considered, the tradeoffs, and the constraints that drove the choice
- C. The team members involved
- D. The implementation timeline
13. Observability for a production agent should capture which of the following to make incidents diagnosable?
- A. Final outputs only
- C. Traces of tool calls, retrieved context, and decision points alongside outputs
- B. Aggregate latency metrics only
- D. User feedback only
14. A partner team wants to adopt your Claude Code setup. What makes adoption durable rather than a one-time handoff?
- A. A recorded walkthrough
- D. Version-controlled shared configuration, common skills and commands, and an onboarding path that new engineers follow
- B. A shared chat channel
- C. Individual coaching sessions
15. Which change best reduces risk in an agent that can both read and modify production records?
- B. Logging modifications for later review
- A. Separating read and write capabilities, granting write access only where the role requires it, and gating writes behind explicit approval
- C. Requiring a longer system prompt about caution
- D. Running the agent during business hours only
Answer key
| # | Answer | Domain | Why |
|---|---|---|---|
| 1 | C | Solution Design & Architecture | Frequently changing knowledge belongs in retrieval, not in weights or a context window |
| 2 | D | Integration | Degradation following a re-index points at chunking, embeddings, and index consistency |
| 3 | A | Integration | At large tool counts, scoping and progressive discovery preserve selection quality |
| 4 | B | Evaluation, Testing & Optimization | Subjective quality becomes measurable through rubrics, labeled sets, and mixed methodology |
| 5 | C | Evaluation, Testing & Optimization | Prompt changes are production changes: regression first, then a controlled comparison |
| 6 | D | Governance, Safety & Risk Management | Compliance boundaries are enforced architecturally before data crosses, and audited |
| 7 | A | Governance, Safety & Risk Management | Confidence and consequence routing makes human review affordable and effective |
| 8 | B | Stakeholder Communication & Lifecycle Management | Honest segmentation plus method and end-to-end outcome is what earns and keeps trust |
| 9 | C | Governance, Safety & Risk Management | GDPR drives minimization, purpose limitation, and erasure, all of which shape design |
| 10 | D | Solution Design & Architecture | Known path, deterministic orchestration; autonomy earns its complexity only when the path is not known |
| 11 | A | Claude Models, Prompting & Context Engineering | A stable cached prefix cuts time-to-first-token and per-request cost together |
| 12 | B | Stakeholder Communication & Lifecycle Management | A decision record captures alternatives and tradeoffs, not just the outcome |
| 13 | C | Evaluation, Testing & Optimization | Without traces of tool calls and retrieved context, failures cannot be localized to a layer |
| 14 | D | Developer Productivity & Operational Enablement | Shared, versioned configuration with an onboarding path is what makes adoption stick |
| 15 | A | Integration | Least privilege plus approval gates on writes reduces the surface rather than monitoring it |
Interpreting your score
| Raw score | Reading |
|---|---|
| 13 to 15 | Comfortable. Review misses and book the exam |
| 10 to 12 | Close. Target weak domains, and do not neglect governance and stakeholder material, which together are 28% of the real exam |
| 7 to 9 | More study needed. Build or dissect an end-to-end system with retrieval, evaluation, and observability |
| 6 or fewer | Start from the exam guide and the engineering articles in official resources |
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