Claude certification practice questions with answers

Correct answer: Trace each material claim to a page or quoted passage.

A fluent summary can still misstate a number, omit an exception or combine unrelated facts. Check claims that could change the reader's decision against the underlying report. For example, if the summary says a policy applies to every supplier, the reviewer should locate the relevant passage and check whether the report excludes small suppliers or particular regions.

Confidence in the wording gives you no evidence of accuracy. Making the summary shorter changes its length, but does not establish that its claims are supported.

Correct answer: Require human approval of the amount and customer ID.

A refund changes a customer's balance, so the application needs to control the action outside the model's instructions. The approval should identify the customer, refund amount and transaction being authorised. If those details change after approval, the application should require a new approval before issuing the refund.

Allowing any amount gives the model excessive authority. A longer policy prompt may help it propose the right action, but the prompt cannot enforce the payment system's authorisation rules.

Correct answer: Check the source owner, version and review date.

First establish whether the file contains prices that are still valid. A document may be well formatted and accurately quoted while being commercially out of date. Identify who maintains the price list, which version was supplied, and whether a newer approved version supersedes it.

Changing prompt adjectives or bullet formatting cannot update the underlying prices. Once you have the correct source, repeat the task and confirm that the output uses that version.

Correct answer: Return a structured error and allow a bounded repair.

The application should explain which required field is missing so the next attempt can correct it. For example, a ticket lookup that requires a ticket ID should report that missing field instead of guessing an ID from an unrelated part of the conversation. Limit the number of repair attempts and ask for clarification if the required information remains unavailable.

An invented default can select the wrong record. Repeating the same invalid call indefinitely consumes time and resources without resolving the missing input.

Correct answer: Treat retrieved content as untrusted and restrict tools.

An email is evidence to inspect, including when it contains text that looks like an instruction. A supplier cannot grant the assistant permission to export customer records by writing that request in an email. The application should restrict accessible tools and data independently of the retrieved text, so a misleading instruction cannot create new authority.

Increasing temperature changes response variability, not the trustworthiness of an email. Promoting the email into privileged instructions makes the trust boundary less clear.

Correct answer: Run a fixed evaluation set against the current version.

Compare the changed prompt with the version currently in use on the same representative invoices. Check the fields the workflow depends on, such as invoice number, currency, line totals and handling of missing values. Include difficult cases so an improvement on a clean invoice does not hide a regression on a scanned or unusual document.

One developer's impression is too narrow to establish reliability. Releasing the change to every invoice before comparing results exposes the whole workflow to an unmeasured change.

Correct answer: Business validation against the account system.

A schema can check that account_number is present and has the required format. The authoritative account system must establish whether that account exists and whether the current user is allowed to act on it. Both checks are needed before the application performs a consequential action.

A larger context window does not create an authorisation check. Schema validation alone cannot establish facts that depend on the live account record.

Correct answer: Authorised legal review against requirements.

The responsible reviewer should compare the clause with the agreement's requirements and decide whether to accept, revise or reject it. The review needs to consider the obligation created by the wording, rather than only whether the prose sounds professional. A draft can be useful while still requiring changes before it becomes part of an agreement.

Publishing the first answer skips that decision. The model's own confidence cannot replace the judgement of the person authorised to approve the clause.

Correct answer: Load a task-specific allowlist.

Give the agent the tools needed for the current task. For a request to summarise ticket status, that might mean allowing ticket lookup and search while withholding ticket deletion or customer-refund tools. This reduces unnecessary capabilities and makes the available choices easier to interpret.

Renaming the tools does not limit what they can do. Exposing all tools leaves unrelated actions available even when the task provides no reason to use them. The application still needs authorisation checks for each tool it permits.

Correct answer: A trace of inputs, evidence, calls, approvals and model version.

A useful trace lets a reviewer reconstruct the decision: which order was considered, which evidence was used, which tools ran, and which approval allowed the action. Recording the model version helps the team investigate whether a later change affected behaviour. The trace should retain the information needed for review under the organisation's data-handling rules.

The final answer alone hides the path that produced the approval. A token count measures usage, but cannot explain why the order was accepted.