22 Prompting, Task Execution & Troubleshooting Practice Questions & Answers

1. 1. A marketing associate types a single line into Claude: "Write something about our new product." The draft that comes back is fluent but generic — it could describe almost any product. What is the most effective single improvement to the prompt?

   • A.State the audience, the purpose, the key product facts, the required format and the target length.✓ Answer
   • B.Ask Claude to rate how good its own draft is and revise anything it scores low.
   • C.Add "make it high-quality, engaging and professional" to the end of the prompt.
   • D.Run the same prompt several times and pick the best of the drafts.

   Generic output is the predictable result of an underspecified request: with no audience, purpose, source facts, format or length, the model must average across every plausible interpretation. Naming those five elements narrows the space of acceptable answers, whereas vague quality adjectives like "engaging" add no information the model can act on.

   Source: Anthropic Claude certification (Associate – Foundations), Domain 1 (Prompting and Task Execution): anatomy of an effective prompt — role, context, task, constraints, output formatReport a problem with this question

2. 2. An associate asks Claude: "Summarize the key obligations in our Q3 vendor agreement." The agreement is not attached or pasted. The response is confidently worded and contains specific-sounding clause references. What should the associate do first?

   • A.Rephrase the request more forcefully so the task is treated as important.
   • B.Ask Claude whether it is familiar with the agreement before requesting the summary again.
   • C.Add "only state facts that are actually in the agreement" to the same prompt and re-send it.
   • D.Paste or attach the actual text of the agreement into the prompt and then ask for the summary.✓ Answer

   A model cannot summarize a document it was never given; anything it produces is reconstructed from generic patterns, which is exactly why the clause numbers look plausible. The fix is structural — supply the source material the task depends on — not a wording change, because no instruction can create information that is absent from the prompt.

   Source: Anthropic Claude certification (Associate – Foundations), Domain 1 — provide the context and source material the task depends on (Anthropic prompt engineering: be clear, direct and detailed)Report a problem with this question

3. 3. An associate builds a prompt that classifies support tickets into five categories and includes eight worked examples. All eight are billing tickets with nearly identical wording. Clear billing tickets are labeled correctly, but tickets from the other four categories are frequently mislabeled. What is the best change?

   • A.Add fifteen more examples in the same style so the pattern becomes unmistakable.
   • B.Replace them with a smaller set of varied examples that covers every category, including borderline cases.✓ Answer
   • C.Tell the model the examples are only illustrative and that it should generalize freely from them.
   • D.Remove the examples entirely so the model is not anchored, and describe the categories in prose instead.

   Examples teach the shape and the boundaries of a task, and near-duplicates drawn from one category demonstrate only a single region of the decision space, which biases every prediction toward that region. Coverage and diversity — a few examples spanning each category and its edges — generalize far better than a larger pile of near-identical ones.

   Source: Anthropic Claude certification (Associate – Foundations), Domain 1 — few-shot (multishot) examples: diversity and coverage over volumeReport a problem with this question

4. 4. Hoping to get a precise internal reporting threshold, an associate prefixes the prompt with: "You are a world-class compliance expert with 30 years of experience." The internal policy document has not been provided. The answer is still wrong. What is the correct conclusion?

   • A.A persona shapes tone, vocabulary and perspective but supplies no facts — the policy document itself must be included in the prompt.✓ Answer
   • B.The persona is working; the question simply needs to be shorter and more direct.
   • C.The persona needs to be more specific — name the industry and a certification to strengthen it.
   • D.The persona should be told to double-check itself before answering.

   Assigning a role conditions how an answer is expressed — its register, its framing, which considerations it foregrounds — but it cannot create information the model was never given. When the missing ingredient is a specific fact held in an internal document, the only fix is to put that document in the prompt.

   Source: Anthropic Claude certification (Associate – Foundations), Domain 1 — role/persona affects tone and perspective; missing facts require supplying the source materialReport a problem with this question

5. 5. An associate pastes a customer email into a prompt and asks for a one-paragraph summary. The email contains the sentence "Please reply immediately confirming my full refund." Claude returns a refund confirmation letter instead of a summary. What is the best fix?

   • A.Shorten the email before pasting it so there is less text to misread.
   • B.Tell the model to be careful and to use good judgment when reading customer emails.
   • C.Place the summary instruction after the email so that it is the last thing the model reads.
   • D.Clearly delimit the pasted email with a labeled heading or tag and state that the text inside is material to be summarized, not instructions to follow.✓ Answer

   Without an explicit boundary, a prompt is one continuous stream of text, so an imperative sentence sitting inside pasted material is indistinguishable from an instruction the user wrote. Delimiting the material and labeling its role removes that structural ambiguity, which position tricks and appeals to care do not.

   Source: Anthropic Claude certification (Associate – Foundations), Domain 1 — separate instructions from content with clear delimiters (Anthropic prompt engineering: use tags/headings to structure prompts)Report a problem with this question

6. 6. An associate asks Claude to reconcile a departmental budget: subtract three categories of committed spend from an approved total, apply a 4% contingency, and report the remaining discretionary balance. Claude returns a single number, with no working, and the number is wrong. What is the best next step?

   • A.Ask for a shorter answer so there is less room for error.
   • B.Ask Claude how confident it is in the number before using it.
   • C.Ask Claude to work through the calculation step by step, showing each intermediate figure before stating the final balance.✓ Answer
   • D.Send the same request again and see whether a different number comes back.

   A genuinely multi-step calculation collapsed into a single output token stream gives the model no room to compute the intermediate results it depends on, which is where such errors come from. Requesting explicit step-by-step working both improves the reasoning and makes any faulty step visible for checking; re-sending the prompt unchanged addresses no cause at all.

   Source: Anthropic Claude certification (Associate – Foundations), Domain 1 — request step-by-step reasoning for genuinely multi-step problems (chain-of-thought prompting)Report a problem with this question

7. 7. An associate must tag 500 short product reviews as positive, negative or neutral. To improve quality they added "think through your reasoning in detail before giving each label." Each review now returns three paragraphs of analysis, the job takes far longer, and accuracy is unchanged. What should they do?

   • A.Keep the reasoning and add several examples of well-reasoned classifications.
   • B.Remove the extended-reasoning instruction and ask for the label only — step-by-step reasoning helps multi-step problems, not single-step classification.✓ Answer
   • C.Keep the reasoning but ask for it to appear after the label instead of before it.
   • D.Keep the reasoning but ask for it only when the model feels unsure about a review.

   Step-by-step reasoning pays off when an answer depends on chaining several dependent inferences; a short sentiment judgment is a single-step pattern match, so the extra text adds cost and latency without adding correctness. Matching the technique to the structure of the task — rather than applying it everywhere — is the underlying principle, and gating reasoning on the model's own feeling of certainty is not a reliable control.

   Source: Anthropic Claude certification (Associate – Foundations), Domain 1 — adapt prompting strategy to task type; reserve extended reasoning for multi-step problemsReport a problem with this question

8. 8. An associate's prompt for a customer-facing announcement reads: "Don't be too formal. Don't use jargon. Don't make it long. Don't use bullet points." The drafts still miss the mark. What is the best revision?

   • A.Repeat each prohibition twice and put the key words in capitals.
   • B.Ask the model to repeat the rules back before it starts writing.
   • C.Rewrite the constraints as a positive specification: a conversational tone, plain language, under 150 words, in continuous prose.✓ Answer
   • D.Add further prohibitions to cover each remaining problem as it appears.

   A prohibition eliminates one option but leaves the entire remaining space open, so a stack of "don'ts" can be fully obeyed and still produce something you did not want. Stating the target positively — the tone, register, length and form you do want — defines a single destination instead of a set of excluded ones.

   Source: Anthropic Claude certification (Associate – Foundations), Domain 1 — state what you want rather than listing what to avoidReport a problem with this question

9. 9. A policy change must be communicated to two groups: the board, who need the financial and risk implications, and warehouse staff, who need to know what changes in their daily routine. A single prompt produced one generic memo that satisfies neither. What is the best approach?

   • A.Keep the memo and ask for a second pass that makes it "less technical".
   • B.Ask for one memo carefully written so that it works for every reader in the company.
   • C.Write two prompts, each naming the reader, what that reader already knows, the decision they must make, and the required tone and length.✓ Answer
   • D.Ask for "a professional version and a simple version" of the same memo.

   Audience determines vocabulary, assumed background, level of detail and what belongs in the first sentence, so an unstated audience forces the model to write for an average reader who does not exist. Naming the reader, their prior knowledge and the decision they face turns each memo into a well-defined task; labels like "simple" or "less technical" leave all of that unspecified.

   Source: Anthropic Claude certification (Associate – Foundations), Domain 1 — specify audience, tone and register explicitly when draftingReport a problem with this question

10. 10. An associate sends one prompt: "Read these 40 survey responses, identify the main themes, write an executive summary, build a slide outline, and draft the announcement email." The response covers everything shallowly and skips the slide outline. What is the best correction?

    • A.Ask for a substantially longer response covering the same five requests.
    • B.Break the work into a sequence of scoped prompts — themes first, then the summary, then the outline, then the email — each building on the checked output of the previous step.✓ Answer
    • C.Add "be thorough and complete, and do not skip any part" and re-send the same prompt.
    • D.Send the five requests as five separate conversations started at the same time.

    A single request carrying five deliverables spreads a fixed response budget across all of them, so each is done shallowly and the least-specified one tends to be dropped. Task decomposition gives every sub-task its own full attention and lets each step consume a verified earlier output, which parallel separate conversations cannot do because they cannot see each other's results.

    Source: Anthropic Claude certification (Associate – Foundations), Domain 1 — task decomposition: break a large request into a sequence of scoped sub-promptsReport a problem with this question

11. 11. An output had the right facts but the wrong structure. In response the associate rewrote the whole prompt at once — new persona, new tone, a length limit, a different format and three new examples. The new output is different but still not usable. What is the best practice going forward?

    • A.Start again from a general-purpose prompt template found online.
    • B.Regenerate the new prompt a few times and keep the best result.
    • C.Return to the original prompt and change only the one thing that was actually wrong — the output-format specification.✓ Answer
    • D.Continue changing several variables at once until some combination happens to work.

    Iterative refinement works by diagnosing the specific defect and changing one variable so that the effect of the change is attributable. Wholesale rewrites confound every change with every other and often break the parts of the prompt that were already producing correct content.

    Source: Anthropic Claude certification (Associate – Foundations), Domain 1 — iterative refinement: diagnose the defect and change one variable at a timeReport a problem with this question

12. 12. After several exchanges in which the associate supplied background documents and agreed on an outline, Claude produces a draft that is about 80% right: two paragraphs are too blunt for the audience, and the risks section is too thin. What is the most efficient next action?

    • A.Regenerate the same request repeatedly until a version comes back with both problems solved.
    • B.Ask for a completely new draft with "softer and more detailed" added to the original request.
    • C.Continue in the same conversation with targeted revision instructions that name the two paragraphs to soften and the section to expand.✓ Answer
    • D.Start a fresh conversation and re-paste all the background with an improved prompt.

    The conversation already carries the background, the agreed outline and the 80% that works, so steering it with specific revision instructions preserves that investment and changes only the identified defects. Restarting discards accumulated context and regenerating leaves the same underspecified request in place, so it only resamples the same failure.

    Source: Anthropic Claude certification (Associate – Foundations), Domain 1 — steer an answer by continuing the conversation rather than restartingReport a problem with this question

13. 13. An associate needs the vendor name, renewal date and annual cost from 30 contracts, to be pasted into a spreadsheet. Claude returns flowing paragraphs that describe each contract differently, and some contracts omit the cost without saying so. What is the best prompt change?

    • A.Ask for bullet points instead of paragraphs.
    • B.Specify a table with named columns — vendor, renewal date, annual cost — one row per contract, and state exactly what to enter when a value is absent.✓ Answer
    • C.Ask for output that is "organized and easy to read".
    • D.Keep the prose and reformat it into the spreadsheet by hand each time.

    When output feeds a downstream system or must be compared field by field, the structure has to be specified explicitly — the columns, the granularity of a row, and the convention for missing values — because otherwise each record is formatted according to whatever the surrounding text suggested. Vague words like "organized" leave the schema undefined, and silent omissions remain invisible without a stated missing-value rule.

    Source: Anthropic Claude certification (Associate – Foundations), Domain 1 — specify an explicit output structure when the output feeds a downstream useReport a problem with this question

14. 14. A classification prompt separates "billing" tickets from "account" tickets. Obvious cases are handled correctly, but tickets that mention both a payment method and a login problem are labeled inconsistently. What is the best improvement?

    • A.Add worked examples of the borderline cases, including counterexamples showing what does not belong in each category, plus an explicit tie-break rule.✓ Answer
    • B.Restate the two category definitions in more formal and precise-sounding language.
    • C.Ask for a confidence score on every ticket and manually review the low-scoring ones.
    • D.Add many more examples of the obvious cases so the categories are better established.

    Definitions describe the centre of a category, but misclassification happens at the boundary, so the boundary itself has to be demonstrated with near-miss pairs and a stated rule for cases that satisfy both definitions. More examples of obvious cases add nothing at the margin, and a self-reported confidence score is not a reliable measure of correctness.

    Source: Anthropic Claude certification (Associate – Foundations), Domain 1 — provide counterexamples and edge cases when a task has tricky boundariesReport a problem with this question

15. 15. A prompt lists 14 requirements, including "cover all nine sections comprehensively", "keep it under 100 words", "maintain a formal register" and "sound warm and conversational". Each draft silently ignores a different subset of the requirements. What is the best fix?

    • A.Resolve the contradictions first, then give a short prioritized set of requirements — or split the work into two prompts with compatible goals.✓ Answer
    • B.Ask the model which requirements it followed, then re-send the prompt.
    • C.Repeat the requirements that were ignored at the end of the prompt.
    • D.Number the requirements 1 to 14 so that none of them can be overlooked.

    "Comprehensive across nine sections" and "under 100 words" cannot both be satisfied, and neither can formal and conversational at the same time, so the model must trade one against another and which one it drops is effectively arbitrary. Removing the conflicts and ranking what matters restores a satisfiable target; numbering or repeating mutually exclusive instructions cannot make them compatible.

    Source: Anthropic Claude certification (Associate – Foundations), Domain 1 — competing or contradictory instructions as a cause of disappointing outputReport a problem with this question

16. 16. For four turns in a row an associate has replied "make it better" to each draft. The drafts change but do not improve. What is the most effective thing to do next?

    • A.Add a persona so the model approaches the draft as an expert would.
    • B.Abandon the thread and start the task over from a blank prompt.
    • C.Describe concretely what a good version must contain and do — the specific criteria it has to meet — or supply an example of an output that would be acceptable.✓ Answer
    • D.Say "make it much better this time" to signal that the previous attempts fell short.

    "Better" specifies no direction of change, so each revision is essentially a random walk around the same content. Stating the criteria a good result must meet, or showing a worked example of one, gives the model a target it can be measured against — which is the same reason explicit success criteria beat vague quality adjectives in the original prompt.

    Source: Anthropic Claude certification (Associate – Foundations), Domain 1 — define what a good result looks like (explicit success criteria or an example)Report a problem with this question

17. 17. An associate needs summaries short enough to fit a ticketing field that allows three sentences. Every prompt so far has said "keep it brief", and the summaries keep arriving at roughly 800 words. What is the best change?

    • A.Ask for the long summary and then ask the model to summarize its own summary.
    • B.State the limit concretely — at most three sentences and no more than 60 words — along with the required form.✓ Answer
    • C.Change "brief" to "very brief" so the instruction carries more weight.
    • D.Trim the extra text by hand before pasting it into the field each time.

    Relative words like "brief" are interpreted against an unknown baseline, so they cannot reliably produce a length that fits a hard downstream limit. A countable constraint is unambiguous and verifiable, and stating it in the prompt is cheaper and more consistent than repairing the output afterwards.

    Source: Anthropic Claude certification (Associate – Foundations), Domain 1 — specify length and format explicitly rather than using relative descriptorsReport a problem with this question

18. 18. An associate pastes twelve months of churn figures into a prompt and writes "Any thoughts?" Claude returns an accurate but unhelpful narrative restating what the numbers show. What is the best improvement?

    • A.Ask for a longer analysis with more detail in it.
    • B.Ask the model to be more insightful and less descriptive next time.
    • C.Ask a specific analytic question — which segments drove the change between Q1 and Q3, and what the data does and does not support — and state the decision the analysis will inform.✓ Answer
    • D.Ask for further thoughts on the same data.

    With no question asked, description is the safest possible response, and length or exhortations to be "insightful" do not change that. An analysis prompt needs the data plus an explicit analytic question and the decision it serves, because those are what define which facts count as relevant.

    Source: Anthropic Claude certification (Associate – Foundations), Domain 1 — adapt strategy by task type: analysis requires the data plus an explicit analytic questionReport a problem with this question

19. 19. An associate asks "Give me some ideas for the spring campaign" and receives five safe, closely related suggestions. They need genuinely different options to take to a planning session. What is the best approach?

    • A.Ask for a stated quantity — say 20 ideas — that must span explicitly different angles such as channel, budget level and audience, then evaluate and narrow them in a follow-up turn.✓ Answer
    • B.Ask the model to pick the single strongest idea straight away and develop it.
    • C.Ask the identical question a second time and merge the two lists.
    • D.Ask for ideas that are "more creative and original" than the first set.

    Brainstorming works in two stages: diverge first, then converge. A small unconstrained request returns the most typical answers, whereas asking for a specified quantity that must vary along named dimensions forces exploration past the obvious, and selection is then a separate, later step rather than something bundled into the same request.

    Source: Anthropic Claude certification (Associate – Foundations), Domain 1 — adapt strategy by task type: brainstorming asks for quantity and divergence first, then convergesReport a problem with this question

20. 20. An associate runs the same extraction prompt on 30 documents. The desired structure is described in a paragraph of prose at the top of the prompt, and the resulting structure varies noticeably from document to document. What is the best change?

    • A.Add "follow the structure exactly as described" to the end of the prompt.
    • B.Ask the model to describe the required structure back before it produces each output.
    • C.Split the work into 30 separate conversations, one per document.
    • D.Include one filled-in example of the exact output template and require every document to be returned in that same form.✓ Answer

    A prose description of a layout leaves interpretive latitude at every ambiguous point, and across many runs those small interpretive choices diverge. A concrete filled-in example removes the ambiguity by showing rather than describing the target form, which is what makes repeated runs consistent enough to compare or process together.

    Source: Anthropic Claude certification (Associate – Foundations), Domain 1 — show the desired output format with an example rather than describing it in proseReport a problem with this question

21. 21. An associate uploads a 20-page internal policy and prompts: "Review this and let me know what you think." The response is a general summary with a few mild suggestions. What the associate actually needed was a list of places where the policy contradicts itself. What is the best revision?

    • A.Add a persona such as "act as a senior policy reviewer" and re-send the same request.
    • B.State the review lens and the deliverable: find statements that contradict one another, quote both passages with their page numbers, and list them in a table.✓ Answer
    • C.Ask for the review to be repeated, this time more critically.
    • D.Ask for a longer and more thorough review of the same document.

    "What do you think" states no criterion, so the model must guess which of dozens of possible review lenses is wanted and defaults to the most generic one. Naming the lens and the deliverable converts an open-ended invitation into a checkable task with a defined output, which neither a persona nor a request for more length can do.

    Source: Anthropic Claude certification (Associate – Foundations), Domain 1 — an ambiguous ask is a primary cause of a disappointing answer; state the task and the deliverableReport a problem with this question

22. 22. Every Monday an associate writes a prompt from memory to produce the same weekly operations report. The wording differs slightly each week and so does the structure of the report, which makes week-to-week comparison difficult. What is the best fix?

    • A.Compare each week's report against the previous one and correct the differences afterwards.
    • B.Tell the model at the end of each session to remember this week's format for future use.
    • C.Ask for more detail each week so that nothing is left out of any report.
    • D.Save a reusable prompt template holding the fixed instructions, constraints and output format, and change only the week's input data.✓ Answer

    Output variation traces back to input variation: if the standing instructions are retyped from memory, they change, and the report changes with them. Holding the fixed requirements constant in a reusable template makes the week's data the only variable, which is what makes successive reports comparable.

    Source: Anthropic Claude certification (Associate – Foundations), Domain 1 — hold standing instructions constant in a reusable prompt; vary only the task inputReport a problem with this question

Practice questions based on the official Claude Certified Associate – Foundations (CCAO-F) exam guide and Anthropic's public documentation. This is an independent study tool, not affiliated with or endorsed by Anthropic, and does not grant certification. The real exam is 60 questions, 120 minutes, passing at a scaled 720/1000, delivered via Pearson VUE ($99). Official certification page →