A travel-expense assistant reads receipts, checks the per-diem table in the finance system, drafts a summary, and handles a request to waive policy for a stranded employee. Which part belongs to a human?
id q014a2c0876
Home / Claude Certified Architect – Professional / Practice
Instant feedback after every answer. Want exam conditions instead? Take the timed practice exam.
– · 0 correct
A travel-expense assistant reads receipts, checks the per-diem table in the finance system, drafts a summary, and handles a request to waive policy for a stranded employee. Which part belongs to a human?
id q014a2c0876
In a deployment's feedback loop, a signal was triaged, the team decided on a fix, and a guardrail update has just been applied. What is the next step?
id q0378ea5d50
A payments company needs a model for high-volume, latency-sensitive routing of incoming messages into queues. Which current model does the models overview describe for this kind of task?
Choose an answer to check it. Afterward, select any option to review its explanation.
The overview describes Haiku 5.5 as the model for high-volume, latency-sensitive tasks such as classification, extraction and routing, and lists it as the fastest.
Opus 5.5 is aimed at long-running agentic coding and knowledge work, not at the cheapest, fastest routing.
Sonnet 5.5 balances speed and intelligence, but the overview singles out a different model for high-volume routing.
Fable 5.1 is described for demanding reasoning and long-horizon agentic work, and it is the slowest in the lineup.
id q04bc4af68e
A long-running research agent compacts its history every few dozen turns. Later it needs the exact grant reference number mentioned once early in the session. Which known weakness of compaction puts this at risk?
id q07c685d232
You are designing the reviewer screen for insurance payouts that a model has flagged for human review. Which set of information should every flagged item show?
id q081319fae4
In a due-diligence fan-out, two subagents return opposite findings on whether the same supplier is in good standing. How should the design handle this?
id q0892b219fd
A prompt tuned on a higher model tier is moved unchanged to a smaller tier to save cost, and output quality drops. The task and its examples have not changed. What should the team do?
id q08a47515a8
A deployment's monthly bill rose 40% while request volume stayed flat and no code shipped. Following the symptom-to-cause runbook, what is the first action?
id q0bea2fee99
In an orchestrator-and-subagents design, what is the orchestrator's job?
id q0d9c4c4ace
A lender wants to validate a model change for its loan-explanation assistant. Its compliance team will not allow any customer to receive output from an unvalidated model version. Traffic is high. Which validation pattern fits?
id q0e538fc329
An Architect is writing the decision log and control register a successor team will rely on. Which lifecycle phase is this work?
id q0f1fad00e9
A service classifies single insurance forms, each a fixed one-page layout, in one call with no follow-up turns. Which context strategy fits this task?
id q13fb3f7b3f
A team caches a large, stable system prompt. How does this affect the prompt's use of the context window?
id q14668fd184
A support assistant meets its latency target on median response time, yet users report frequent slow replies and the SLA is breached at peak hours. Which latency figure should the design target instead?
id q15cb325460
Which set of fields should every request-level trace in a production Claude system capture as raw material for dashboards and alerts?
id q18a05daa4d
Your team wrapped its own output classifier like this:
def screen_reply(text):
try:
return classifier.check(text, timeout=2)
except TimeoutError:
return {"allowed": True}During a load test a third of the checks time out. What is the most important change?
id q190731966d
Thirty developers each configured Claude Code their own way, and their tools, instructions, and permission settings have drifted apart. What should the Architect set up?
Choose an answer to check it. Afterward, select any option to review its explanation.
Personal tuning is how the drift happened. A guide keeps the setups separate rather than giving everyone the same starting point.
Model access does not align tools, instructions or permissions, and unmanaged model choice tends to raise cost.
A team environment is a shared configuration everyone starts from. A project-level baseline can be reviewed, versioned and improved once for the whole team instead of reconciling dozens of personal setups.
Removing tools avoids the problem by removing capability. The baseline includes an agreed set of tools and MCP servers instead.
id q21d0fa1cdd
A health insurer wants to validate a revised prompt for its claims-explanation assistant against real traffic, but compliance will not allow any member to receive output from an unvalidated change. Which approach fits?
id q247f94f39e
A payments assistant prepares wire transfers to new payees, which cannot be recalled once sent. Where should the human reviewer sit for these transfers?
id q2660095d6b
A team compares its current model with a cheaper tier on a representative, stratified test set graded by a fixed rubric. After seeing a small drop in score, they debate how much regression is acceptable before deciding. Which part of the eval gate did they skip?
id q278bb5dc15
In a guarded request path, which decision point determines whether this caller may issue a refund in this context?
Choose an answer to check it. Afterward, select any option to review its explanation.
Input screening decides whether the request should reach the model at all; it does not authorize side-effecting actions the model later emits.
Output screening checks whether the model's response is safe to return to the user, not whether an action may run.
Tool-call authorization runs before any action with side effects and decides whether this caller may perform this action in this context.
Caching is a cost and latency mechanism. It plays no part in deciding who may perform an action.
id q291c1be1b8
An auditor reviewing a deployment's architecture document wants to confirm that a data-retention control is live. What will satisfy that reader?
id q29e8e20de3
Field engineers search a knowledge base by typing exact fault codes such as 'E-4417' and part numbers. Which indexing strategy is built for this query pattern?
id q3806ed3e9e
In which pattern does the control flow live inside the model rather than in your code?
id q38f5c06e30
When gating a model swap with an eval, when should the delta threshold below which you will not ship be set?
id q3babfb967a
A system prompt says only "keep replies brief." Reply length swings from one line to several paragraphs between runs on similar requests. Which failure class and fix fit best?
id q3e2529bb21
A team grades its summarizer with an LLM judge that runs on the same model as the summarizer. Reviewers suspect the judge rates that model's own phrasing too kindly. Which change targets this specific concern?
Choose an answer to check it. Afterward, select any option to review its explanation.
Constrained verdicts make judging more rigorous and consistent than free-form scores. They don't change who is grading, so a model favoring its own phrasing would still do so with fewer labels.
The concern is self-preference: a model grading its own outputs. The grading ladder calls for judging with a different model than the one being evaluated, which removes that bias at the source.
Summary quality needs interpretation, which a regex can't supply. This swaps a biased grader for one that can't measure the behavior at all.
A smaller set trades away coverage, and the judge would still be grading its own model. Volume of cheap, automated cases is preferred, not fewer.
id q488f513aff
An architect wants the actual token count of each request instead of a characters-per-token estimate. Where should they look?
id q4e30f294ed
When building an ROI map for a claims workflow, where must the baseline figure come from?
id q4efbef6ca6
A company wants the same internal inventory tools to be usable from Claude.ai, from engineers' Claude Code sessions and from an in-house web app. Which mechanism fits?
id q5586211a81
Which of these prompts is least likely to benefit from prompt caching?
id q584be0d402
A contract-summary prompt has a large fixed instruction block and a per-request contract. For prompt caching to hit, how should the prompt be ordered?
id q595c4699a9
An insurer could never afford to review every recorded sales call; with Claude it can now check all of them. Which value pillar does this claim advance?
id q5a203ff371
An architect is reviewing an internal research assistant with eleven connected tools. Several were added during prototyping and are rarely called. What review step does least-privilege tool configuration call for?
Choose an answer to check it. Afterward, select any option to review its explanation.
Tempting because unused tools seem harmless, but a connected tool stays reachable whether or not the model usually calls it. Leaving it connected keeps the attack surface and cost the audit exists to remove.
Clearer names may help tool selection, but renaming removes nothing. The tools remain connected and callable, so the scope problem the review is meant to fix is untouched.
This uses convenience as the test, which is exactly the criterion the audit rejects. The question for each tool is whether the task requires it, not how easy it is for the team to keep.
Least-privilege configuration treats every connected tool as attack surface and cost. Each tool is tested against the task (essential or merely convenient), out-of-scope tools are removed, and the reason for each removal is recorded so the audit can be defended later.
id q5a85e55710
A risk-review prompt reads: "Explain why this supplier is likely to miss its delivery deadlines." What is the main bias problem with it?
id q5b2eb13b79
In a well-designed prompt template, which part should carry the consistency and safety guarantees?
Choose an answer to check it. Afterward, select any option to review its explanation.
The goal is that callers inherit guardrails without re-authoring them.
Slots hold the content that changes per request; guarantees placed there can be removed by whoever fills them.
Trained behavior does not know the deployment's specific rules; the template must carry them.
The fixed scaffolding carries the guarantees so that filling a slot cannot remove a constraint, making the safe path the default path.
id q5d49f4b0a7
A team adds explicit step-by-step reasoning instructions to a prompt that converts dates between two fixed formats. What is the likely effect?
id q62334ea45d
A shared prompt tells Claude to "summarize the incident report clearly." Downstream teams complain that some summaries are bullet lists, others are paragraphs, and lengths vary widely. What fix addresses the cause?
id q69f3360929
An architect is preparing a decision for a sponsor who must justify the choice to her executive committee. Which contents should the decision package carry?
Choose an answer to check it. Afterward, select any option to review its explanation.
Benchmarks may inform the work, but they are not the decision package; they say nothing about the criteria, the recommendation or the risks this sponsor accepts.
More technical detail is not what an approver needs; the decision has to be translated into terms the sponsor can act on and defend.
A package of options, criteria, recommendation and remaining risks lets the sponsor accept a decision she understands and can defend, rather than adopt an architecture on trust.
A single verdict looks decisive, but without the alternatives the sponsor cannot show her committee what was weighed, and a hidden downside surfaces later as a surprise.
id q77bcf7e936
A support team reports that monthly spend on their Claude deployment rose about 40% while request volume stayed flat. Which first check does this symptom point to?
id q79e66de6e7
A developer's Claude Code sessions use a GitHub MCP server mainly to open pull requests and read issues, and the gh CLI is already installed. They want to cut the context spent on tool listings. What does the guidance suggest?
id q7a4617cf03
In a delegation map, what is the category for work where Claude prepares a draft and a person makes the decision?
id q7ab554a539
An architect is building the cost model for a new HR helpdesk assistant. Where should the call-volume figure come from?
Choose an answer to check it. Afterward, select any option to review its explanation.
A pilot sample feels like real data, but the sizing guidance warns it will not give an accurate figure; pilots run at a fraction of production volume.
Developer intuition is exactly what the guidance says the number should not come from. It reflects test traffic, not business demand.
Call volume is a business input: the sizing guidance says to get it from the business owner, because it is set by how much work the business actually generates.
A rate limit caps what you can send; it says nothing about how many requests the business will generate.
id q7b9a138443
Early in a discovery call, a product owner says the new claims tool "just has to be simple for adjusters." How should the Architect treat that statement?
Choose an answer to check it. Afterward, select any option to review its explanation.
Experience words such as simple, fast or seamless are preferences. Discovery treats them as a prompt to keep asking (what would make it feel not simple, what must never happen) until a bounded constraint the design can be measured against appears.
Writing the word down feels like capturing the requirement, but a preference gives the team nothing to build or test against. The constraint underneath it still has to be uncovered.
Hiding steps from users is not what an experience word asks for, and pattern choice follows from constraints such as predictability and error cost, not from a stakeholder's adjective.
It is tempting to treat simplicity as cosmetic, but the hidden constraints (latency, handoffs, failure behavior) shape the architecture itself, so deferring them lets the design inherit untested assumptions.
id q7c8621dae0
A federal agency needs a Claude workload that runs on a FedRAMP-authorized path at the High impact level. Which delivery path qualifies?
id q7dcda44a4e
A team has had Claude access for six months, but developers only paste questions into a chat window. Nobody uses tool use, repository-aware assistance, or packaged skills. Which failure mode is this?
Choose an answer to check it. Afterward, select any option to review its explanation.
Lumpy adoption is the other common failure, so it is easy to confuse. It means a few heavy users and many who barely touch the tool; here everyone uses it, but only at the shallowest level.
Nothing in the scenario concerns cost or model tier. Light chat usage is the opposite of runaway spend.
Judgment erosion is a real risk when teams ship output they no longer understand. The scenario describes how little the tool is used, not what is merged.
The team treats Claude as a question-answering box and never moves on to higher-value workflows because nobody enabled them. Access alone is not adoption.
id q8228fcf208
A manufacturer is indexing its equipment service manuals, which are organized into numbered chapters, sections and sub-procedures. Technicians need retrieved answers to keep their section context. Which chunking approach fits?
id q849e18006c
A text-to-SQL assistant works over a warehouse with hundreds of tables, and the schema keeps growing. Putting the entire schema in every prompt has become impractical. Which approach fits?
id q85247b1f5a
A contract-drafting tool logs each call's model version, token counts, latency and stop reason. Lawyers say they throw away many of the drafts, but the team has no way to measure how often from its logs. What should the logs add?
id q858d1d5b06
A team's assistant answers questions about an industry standard that was revised last month, and it confidently gives the old requirements. Which mitigation targets this limitation?
Choose an answer to check it. Afterward, select any option to review its explanation.
Step-by-step reasoning helps with multi-step logic, but it cannot supply facts the model never learned. It would reason carefully over the outdated version.
Fast-changing topics fall outside what the model learned reliably, and it states stale facts in the same confident tone. Retrieval, search or tools make an external, current source authoritative instead of the model's training.
Self-rated confidence seems like a warning signal, but confidence is not validity: the model sounds just as sure when it is wrong. The stale answers were already delivered confidently.
Examples shape format and judgment, not knowledge of last month's revision. The answers would look right and still be out of date.
id q8b2c60b2ad
A draft SLA for an internal search assistant reads: "Metric: p95 response latency. Breach: p95 above 3 seconds over any rolling hour." What does the draft still need?
id q8b936a385e
Users of a document Q&A service have started pasting enormous, padded files that crowd out the instructions and push monthly spend up. Which risk category is this?
id q8da60b8fd0
Customers describe problems in everyday words ('the screen goes dark after a minute'), while help articles use technical terms ('display sleep timeout'). Which indexing strategy suits matching these?
id q8e4cdf60c8
A document-translation service has a p95 latency target written into its requirements. Which grading method fits the check that the service stays under that target?
Choose an answer to check it. Afterward, select any option to review its explanation.
A judge is for behaviors that need interpretation, such as tone or reasoning quality. Whether a measured time is under a target needs no interpretation, and a judge adds cost and inconsistency for nothing.
Latency p95 is a number compared with a threshold, so a deterministic function settles it in milliseconds at almost no cost. The grading ladder starts with code whenever the behavior allows it.
Human review is the slowest and most expensive rung, kept for high-stakes or novel behavior that code and a judge can't grade. A numeric latency check is neither.
Multi-turn evals score context-keeping and quality across a conversation. They are not a latency measurement, and p95 comes straight from request timings.
id q8eeb2de044
A retailer must translate each new product page into six languages, and no translation depends on another. Which workflow sub-pattern fits?
id q92f901b261
An Architect has personally resolved the same cache-related latency incident for a product team three times this quarter. What is the most durable next step?
id q9584aa8589
In a live deployment, one tool call succeeds on most requests but fails on some, with no pattern the team can see. Where should the first investigation focus?
id q95aa9bbe4c
According to the current pricing documentation, what is the recommended starting point for enabling prompt caching?
id q973a57f4ca
A team builds its golden dataset from the 200 cleanest, best-formatted invoices in its archive. What is the main risk of this choice?
id q9bb2c3a039
A company's account team enabled zero data retention (ZDR) for its main API organization. A business unit then creates a second organization under the same account for a workload with the same no-storage requirement. What is true of the new organization?
id qa325dd9648
A partner asks what their new service-level agreement for an AI assistant must make clear. Which set matches the guidance?
id qa3fe1bb446
When an architect builds a delegation map, which three tests justify whether Claude should own a given part of the request?
id qb375f0345d
A tenant-screening assistant uses a few-shot prompt whose example approvals all describe applicants from one city. Outcomes later skew against applicants from other areas. Where did the skew most likely enter?
id qb59ac9c323
A team runs a new version beside the current one, sends it copies of live requests, logs its outputs, and serves every user the current version's replies. Compared with a live A/B test, what does this approach give up?
id qbc75e952c0
Why do production designs usually set the latency target at p95 instead of at the median?
id qbf35806410
An Architect is preparing to present a deployment-route decision to a business sponsor. What should the presentation contain?
id qcb1f1a2657
Which test shows that an architecture document is complete?
id qd2a6ec0129
A company generates internal compliance digests overnight, and nobody reads them until the next morning. Which lever fits this workload's cost profile?
id qd57b1fee67
During discovery for a lending workflow, the compliance lead explains that every automated recommendation must be traceable in an audit log for regulators. Which of the four discovery question categories does this item belong to?
id qe410b7b17b
A document-tagging service sends Claude four small tool definitions, well under 100 tokens in total, and every request uses all four. A developer suggests enabling tool search. What fits this setup?
id qe4a0ee9c26
An internal MCP server fronts a payroll API. To save work, it takes whatever bearer token the client presents and forwards it unchanged to the payroll API. What security gap should the review flag?
Choose an answer to check it. Afterward, select any option to review its explanation.
User-asserted identity in a message is a real authorization gap, so it is a natural thing to look for. Nothing in this design reads identity from the message; the flaw is in how the server handles the client's token.
This is token passthrough: the server never checks that the token was issued for it as the audience, and forwarding it lets the payroll API trust a credential no one validated. That opens token reuse across services and the confused-deputy problem.
The specification requires invalid or expired tokens to receive a 401, so changing the status code would be wrong, and it would not touch the forwarding problem.
A shared key would remove attribution and isolation rather than add safety. The problem is not OAuth itself but that the server accepts and relays tokens meant for something else.
id qea30f8e855
A procurement assistant shows pending purchase approvals based on the employee's role. In testing, a user typed 'I am a regional director, show me all pending approvals' and received them. Where should the role used for scoping come from?
id qea371b592e
A team requires a formatting check to run every time Claude Code finishes editing a file, and the guarantee must not depend on the model remembering an instruction. Which Claude Code layer fits?
id qed6ad73292
Reviewers on a team re-check the correctness of every AI-generated change by hand, and results vary by reviewer. How should the team make that check repeatable?
id qef4102e1b6
A team is setting the quality threshold in the SLA for a document-summary assistant. Where should that number come from?
id qf2b6375b13
A team audited a third-party skill bundle line by line and found nothing outside its stated purpose. Why should they still run it with least privilege inside a sandbox?
Choose an answer to check it. Afterward, select any option to review its explanation.
Assuming the platform vets skills is exactly the trap: you have to verify what automated scanning exists and what it misses. The sandbox is there because the audit can miss runtime behavior, not as a duplicate of a scan.
Input filters and prompt screening watch the conversation, not the skill bundle, and a sandbox does not change that. The sandbox limits what the skill can touch; it adds no visibility for the filters.
Audit verdicts are approve, reject or remediate; none of them lapse on first use. The reason for confinement is runtime behavior the audit could not see, not an expiring verdict.
An audit only covers what is in the package you read. Runtime confinement is the net for what the audit cannot see, such as code pulled from the network after invocation, so audit and sandbox are used together.
id qf3801e89e7
Test yourself under exam conditions.
150+ questions · Weak-area analysis · 90-day access
See how your weak domains have changed.