8 of the 15 questions in this set, with the correct answer marked and every option explained.
1. Claude produces a market summary containing a specific statistic and a company name. Before including it in a client deliverable, what should the user do?
✓Verify the statistic against a primary source, because specific figures and names are exactly the details a language model can state fluently but incorrectly.
Confident phrasing carries no guarantee of accuracy, and precise numbers and proper nouns are the highest-risk category. Verification against the underlying source is the professional standard for anything going to a client.
✗Include it as written, since the model would have flagged uncertainty if the figure were unreliable.
Models do not reliably signal uncertainty about specific facts. Fluency and confidence are properties of the writing, not evidence of correctness.
✗Ask Claude whether the statistic is correct and accept its answer.
Asking the same model to check itself does not provide independent verification. It may well restate the error with equal confidence.
✗Rephrase the sentence so the statistic sounds less precise.
This disguises the problem rather than resolving it, and vague sourcing in a client deliverable is its own failure.
2. Which type of task carries the highest risk of an unverified AI-generated error causing real harm?
✓Producing figures or citations for a regulatory filing or legal document.
These combine two risk factors: outputs that must be factually exact, and consequences — legal, financial, regulatory — that are severe and hard to reverse. This is where human verification is least optional.
✗Brainstorming candidate names for an internal project.
Low stakes and inherently subjective. A poor suggestion is simply discarded at no cost.
✗Rewriting an internal email to sound more concise.
The author reads and sends it, so review is naturally built into the workflow, and the consequences of an awkward sentence are minor.
✗Generating discussion questions for a team retrospective.
Open-ended and non-factual. There is no correct answer to get wrong.
3. A user asks Claude to summarize a 40-page document they have uploaded. Which check best validates the summary?
✓Spot-check several specific claims in the summary against the corresponding sections of the source document.
Tracing claims back to the source is what tests faithfulness. It catches both fabricated detail and, just as importantly, material the summary quietly omitted or distorted.
✗Ask Claude to rate its own summary out of ten.
Self-rating is not independent evidence. A model can score an inaccurate summary highly.
✗Check that the summary is shorter than the original.
Brevity is a formatting property. A short summary can be entirely wrong.
✗Confirm the summary uses the same vocabulary as the source.
Shared vocabulary is weak evidence — a summary can reuse terminology while misstating the relationships between those terms.
4. Which prompt is most likely to produce a directly usable result?
✓"Write a 150-word update for our operations team explaining the Q3 shipping delay, in a factual tone, ending with the three actions we are taking."
It specifies audience, length, topic, tone and required structure. Each of those removes a decision the model would otherwise have to guess at, which is why constrained prompts need less rework.
✗"Write something about the shipping situation."
Audience, length, tone, purpose and content are all unspecified, so the output is a guess that will very likely need rewriting.
✗"Shipping delay."
A fragment with no instruction at all. The model cannot know whether you want an explanation, an email, an analysis or a definition.
✗"Do your best on the shipping topic and make it good."
"Good" is not a specification. It conveys a desire for quality without any of the criteria by which quality would be judged.
5. A user is dissatisfied with a first draft. What is the most effective next step?
✓Give specific feedback about what to change — tone, length, emphasis, structure — and ask for a revision.
Conversation is iterative, and targeted feedback preserves what worked while fixing what did not. It is nearly always faster and better than starting over.
✗Start a completely new conversation with the same prompt and hope for a better sample.
Resampling the same prompt discards the context of what was wrong and relies on luck. The underlying under-specification remains.
✗Ask the same question repeatedly until the wording improves.
Repetition without new guidance provides no information about what 'better' means.
✗Accept the draft and edit it manually without telling Claude what was wrong.
Sometimes reasonable for a small fix, but it forgoes the chance to get a better draft and teaches the model nothing about your preferences for the rest of the session.
6. Why does providing an example of the desired output format usually improve results?
✓An example communicates structure, tone and level of detail more precisely than a description of them, leaving less room for interpretation.
Describing a format in words is lossy — 'professional but friendly' means different things to different readers. An example collapses that ambiguity by demonstrating exactly what is wanted.
✗Examples cause the model to retrieve the source document the example came from.
Nothing is retrieved. The example simply appears in context as a demonstration to follow.
✗Examples reduce the cost of the request.
Examples add tokens, so they modestly increase cost. They are worth it because they reduce rework, not because they are cheaper.
✗Examples permanently change how the model responds to all future users.
An example in a conversation affects that conversation only. It does not modify the model.
7. A business user wants Claude to work with the same set of reference documents across many separate conversations without re-uploading them each time. Which capability fits?
✓A Project, which holds shared knowledge and instructions that every conversation inside it can draw on.
Projects exist precisely to give a set of conversations a shared, persistent body of context. Uploading the reference material once makes it available to every chat in that project.
✗Increasing the model's context window setting.
Window size governs how much fits in a single conversation. It does not carry material between separate conversations.
✗Pasting the documents into each new conversation.
This works but is exactly the repetitive effort the question asks to avoid, and it invites inconsistency as versions drift.
✗Asking Claude to memorize the documents in a first conversation.
A conversation does not persist into unrelated conversations by default, and the model cannot commit material to memory on request in that way.
8. Which statement about Claude's model tiers is most useful when choosing one for a business task?
✓Larger, more capable models suit complex reasoning and nuanced writing, while smaller, faster models suit high-volume, simpler tasks where speed and cost matter more.
Tier selection is a cost-capability trade-off. Matching the tier to the task's difficulty avoids both overpaying for simple work and under-serving genuinely hard reasoning.
✗The largest model should always be used, since capability is the only consideration.
This ignores latency and cost, which matter a great deal at volume. For classification or simple extraction, a faster model is often the better engineering choice.
✗The smallest model should always be used, since all tiers produce identical output.
They do not produce identical output. Capability differences are real and show up most on complex reasoning and long-horizon tasks.
✗Model tier determines which languages Claude can read.
Multilingual capability is broadly present across tiers rather than being the distinguishing factor between them.
7 more questions in the app
Practise the full 15-question set with a timer, scoring and progress tracking.