Claude certification sample questions, explained

A marketing team's Claude Project has produced accurate campaign briefs for months. This week, every brief omits the regional compliance paragraph. The prompt template has not changed. On Monday the team uploaded forty new files to Project knowledge, including three archived brief templates that predate the compliance requirement. What is the BEST first step?

• A. Rewrite the prompt to state that the compliance paragraph is mandatory.
• B. Put the Project on a more capable model, reasoning that its knowledge base is now much bigger.
• C. Identify which layer actually changed (the knowledge base, not the prompt) and remove or supersede the outdated templates before altering anything else.
• D. Add a reviewer who checks each brief for the missing paragraph before it ships.

Answer: C

The controlling fact is "the prompt template has not changed", beside a knowledge base that changed on Monday. A is the most tempting miss: it repairs the loudest variable rather than the broken one, and leaves three contradictory templates to resurface. B is premature escalation: nothing suggests a capability limit, only a source-of-truth conflict. D catches the symptom every time and the cause never.

A finance analyst wants Claude to draft supplier payment-risk notes using a connector that already has read access to the shared finance drive. That drive also holds payroll files the analyst is not authorised to use for this purpose. Which TWO steps matter most before the workflow runs? Select TWO.

• A. Confirm that the purpose is authorised for this data, separately from the connection being authorised.
• B. Rely on the connector's existing permissions, which IT approved when it was set up.
• C. Use a data source or account limited to the approved supplier folders, then test that payroll files cannot be retrieved.
• D. Add a line to every output noting that it was drafted with AI assistance.
• E. Instruct Claude to confirm that it will not open the payroll files.

Answer: A and C

Judge all five independently before counting. B is access mistaken for authorisation: a connection approved for one purpose is not permission for another. D substitutes disclosure for control: telling the reader an AI helped does nothing about payroll. E is the self-certification trap: an assurance produced by the system being constrained is not a constraint. That leaves exactly two: authorised purpose, and an access boundary enforced outside the prompt.

An HR operations lead has spent one long Claude conversation drafting a forty-page onboarding handbook. Late in the session, Claude begins restating policies that were corrected an hour earlier and misnames the company's leave scheme. Two chapters remain and are due today. What is the MOST appropriate next step?

• A. Ask Claude to try harder and re-read the earlier messages before each answer.
• B. Switch the same conversation to the most capable model, reasoning that a stronger model will remember more.
• C. Have Claude write a short summary of the agreed decisions and corrected facts, verify it, then start a fresh conversation or a Project that carries the summary as knowledge for the remaining chapters.
• D. Paste the entire transcript back into the conversation so that nothing is lost.

Answer: C

The controlling fact is a very long session that has begun to drift, which the Associate blueprint treats as a context-management decision: restart, summarise or persist. A treats a capacity limit as an effort problem; an instruction cannot restore what the context no longer holds reliably. B escalates capability without diagnosing the layer that failed, and the drift travels with the conversation. D adds the most material to the session that is already overloaded. C preserves the verified decisions in a compact form and gives the remaining work a clean context.

A regional bank's head of operations asks the architect for a written commitment that a new Claude-based document triage service will classify 99.9 percent of incoming forms correctly from launch day. A four-week pilot on 1,200 forms measured 96.4 percent, with most misses concentrated in one handwritten form type. The launch date is fixed. Which response BEST manages the expectation?

• A. Accept the 99.9 percent figure in the statement of work to keep the relationship positive, and plan to improve accuracy after launch.
• B. Present the pilot measurement by form type, propose a service level tied to the measured baseline with human review routed for the handwritten type, and agree how the target is re-baselined after launch.
• C. Decline to commit to any accuracy figure, on the grounds that language model output cannot be guaranteed.
• D. Move the service to a larger model tier before the meeting so that the higher number can be promised.

Answer: B

The Professional blueprint's stakeholder domain asks for expectation alignment including service levels, and the controlling facts are a measured 96.4 percent and a fixed date. A agrees to a figure nothing has measured; a commitment without evidence becomes a breach on day one. C walks away from the conversation the objective requires, when a defensible number exists. D is capability escalation without an evaluation; nothing in the pilot says a larger tier reads handwriting better. B anchors the commitment to evidence, contains the known weak type with a control, and keeps the number revisable.

A logistics startup's invoice-parsing service calls Claude through the SDK using a model alias that always resolves to the newest release. After a release, a downstream validator starts rejecting eight percent of outputs because a date field now arrives in a different format. Nothing in the service's own code changed. Which change BEST prevents a repeat?

• A. Add a retry that resubmits any rejected output until the validator accepts it.
• B. Pin the service to a specific model version, state the date format in the output schema or prompt constraints, and run the evaluation set against a new version before promoting it.
• C. Loosen the validator so that it accepts any date format.
• D. Move the workload to the Message Batches API so that failures cost less.

Answer: B

Two published skills meet here: configuration management, which names model version pinning, and output handling, which names response validation. A retries a systematic change, so every retry fails the same way and the bill grows. C removes the control that caught the problem; the systems downstream still need one format. D changes cost and latency tolerance, not the stability of the output. B makes release a decision the team takes, constrains the field, and tests before promoting.

A support agent built on the Claude Agent SDK sometimes stops before a refund completes, because the loop ends whenever the assistant's text contains the word "done". On other runs it keeps going after the work is finished until a fixed cap of eight iterations halts it. Which change BEST fixes the loop?

• A. Raise the iteration cap to twenty so that the agent has more room to finish.
• B. Continue while the response's stop reason indicates a tool call and end when the model signals the turn is over, appending each tool result to the conversation before the next request.
• C. Instruct the model in the system prompt never to say "done" until every tool call has completed.
• D. Treat the presence of any assistant text in a response as the completion signal.

Answer: B

The Architect Foundations guide names the anti-patterns directly: parsing natural language for termination, an arbitrary iteration cap as the primary stop, and treating assistant text as completion. A keeps the cap as the mechanism and leaves the early stop untouched. Capplies prompt guidance to a control-flow problem and still reads a word to decide. D is the third named anti-pattern; text and tool calls can arrive together. B drives the loop from the structured signal the API returns and keeps the history the next step reasons over.