Practice — CCAO-F Domain 1: Prompting and Task Execution (14%)
13 scenario-based MCQs. Answer key + explanations at the bottom.
Q1
An operations manager asks Claude to "write an email to our vendor about the delayed shipment." The draft is polite but generic — it doesn't reference the contract terms or the firm-but-collaborative tone the company uses with this vendor. Which revision to the prompt would most improve the next draft?
A. Ask Claude to regenerate the draft several times and choose the strongest of the resulting versions B. Switch the conversation to the most capable model available before asking again C. Add the relevant context (contract terms, relationship history) plus explicit tone and length constraints D. Tell Claude to make the email "more specific and less generic" than the previous version
Q2
The AI Fluency framework describes day-to-day prompt iteration as the Description–Discernment loop, in which "every prompt is a hypothesis." What does this mean in practice?
A. You should draft several alternative prompts and compare them before sending any of them B. Each output should be evaluated for what it reveals about the prompt, and the description revised accordingly C. Claude internally tests multiple interpretations of the prompt and returns the most likely one D. Prompts should be kept short enough to test quickly, then expanded once they start working
Q3
A product manager needs a competitive analysis covering six competitors, drawing on a stack of uploaded research notes. A single prompt asking for the complete document produced a shallow, uneven result. Which approach best applies task decomposition?
A. Request an outline first, then generate and review one section at a time, then ask for a final synthesis B. Resend the same request with far more detailed instructions covering every expected section of the document C. Ask Claude to think carefully and take as long as it needs before producing the full document D. Move the notes into six separate conversations, one per competitor, and merge the outputs by hand
Q4
A marketing associate types "write a launch post for our new feature" with no other information, and Claude produces fluent but generic copy that could describe almost any product. What best explains the result?
A. Claude's knowledge cutoff prevents it from writing about products launched recently B. Claude retrieved inaccurate details about the company from the web while drafting C. Claude located the most similar product launch in its training data and paraphrased that company's post D. Claude had no product, audience, or voice details to steer it, so it generated plausible generic text
Q5
An analyst uses Claude each week to turn meeting notes into a status update that gets pasted into a shared spreadsheet. Some weeks the output arrives as paragraphs, other weeks as bullet points. Which prompt change most reliably fixes this?
A. End every request with a reminder that formatting consistency matters to the whole team B. Specify the exact output format in the prompt — a table with named columns in a fixed order — and include one example row C. Ask Claude to choose whichever format best fits the content of that week's notes D. Add the phrase "use structured output" so Claude understands that prose paragraphs are not acceptable
Q6
For which situation does breaking a request into sequenced smaller prompts clearly beat writing one longer, more detailed prompt?
A. Whenever the finished deliverable will run longer than a page or two of text B. Whenever the task draws on more than one uploaded source document C. When later steps depend on earlier results and you want to review and steer intermediate output before continuing D. When the output contains errors, since shorter prompts produce proportionally fewer mistakes
Q7
A consultant pastes a 90-page report into a conversation, preceded by a list of eight formatting and scope rules. Claude's summary follows some rules and ignores others. The document itself must stay in the prompt. Which prompt-level fix is most likely to help?
A. Ask Claude to confirm that it has read and understood all eight rules before it begins summarizing B. Move the rules into the middle of the document so they sit physically closer to the content they govern C. Rewrite the rules in more formal, contractual language so that they carry more authority D. Restate the most important rules after the document, so they appear near the end of the prompt
Q8
After a disappointing first draft, a teammate clicks "retry" four times with the same prompt, hoping for a better result. Why is this a weak iteration strategy?
A. Without a revised description the outputs won't systematically improve, and the draft's clues about what was under-specified go unused B. Retrying only produces meaningfully different drafts on the most capable models, which sample more diverse responses C. Each retry consumes context window, so the conversation will need to be restarted after a handful of attempts D. Claude remembers each rejected draft in the conversation and steers away from those approaches, progressively narrowing the drafts it can produce
Q9
A brand team wants Claude to draft social copy in their established voice. Prompts describing the voice ("warm, confident, never salesy") keep producing near-misses. Per Anthropic's prompting guidance, what should they add to the prompt?
A. A longer and more precise list of adjectives that pin down the brand voice B. Two or three examples of past on-voice posts for Claude to match C. A role assignment such as "You are our senior brand copywriter" D. An instruction telling Claude to study the company's public website before drafting
Q10
Anthropic's prompting guidance orders its core techniques and advises walking the list top to bottom when a prompt underperforms. Which fix does that order say to try first?
A. State the task, constraints, and expected format directly instead of making Claude infer them B. Add two or three input/output examples showing what a good response looks like C. Ask Claude to reason step by step through the problem before giving its answer D. Assign Claude an expert role that frames how it should approach the response
Q11
An office manager is delegating a Cowork task: cleaning up a shared folder of roughly 200 mixed files. Which request sets the task up best?
A. "Tidy this folder however you see fit — you know what good organization looks like, so use your judgment on structure and naming, and I'll correct anything I don't like afterward." B. "Before touching anything, list every file in the folder so I can classify each one myself first." C. "Organize this folder the way most well-run companies organize their shared drives, following whatever conventions are standard for a business of our size." D. "Sort these files into the four folders I've named below, work on copies rather than originals, flag anything you can't classify, and end with a summary of what you moved."
Q12
An education coordinator asks Claude to "summarize our district's new attendance policy." The policy is an internal document that has never been shared in the conversation, yet Claude returns a confident, detailed summary. Which prompting practice would have prevented this?
A. Asking Claude to double-check its own summary for errors before responding B. Asking the question again in a fresh conversation to clear out stale context C. Uploading the policy document and instructing Claude to summarize only from it D. Asking Claude to include citations so that unsupported claims become visible
Q13
A team lead asks an associate to "get Claude to pull numbers from our internal sales database every morning and email a summary to the team." No connector to that database exists in their workspace. What is the associate's best move?
A. Hand the database-integration piece to the developer team, and prototype the summary-writing prompt on exported sample data B. Keep refining the prompt wording until Claude works out a way to reach the database on its own C. Paste the database credentials into the conversation so Claude can log in and query it directly D. Move the entire request into a Project's custom instructions, since Projects persist across conversations and can run recurring tasks each morning
Answers
Q1: C. The draft is generic because the prompt withheld everything that would make it specific: the contract terms, the relationship history, and the desired tone. Adding context plus explicit constraints is prompt structuring (context, task, format, constraints) working as designed. (D) is the tempting runner-up but fails because "more specific" gives Claude nothing to be specific with — the missing information still isn't in the conversation. (B) is the classic wrong-layer distractor: no model upgrade supplies facts you never provided, and (A) just resamples from the same underspecified prompt.
Q2: B. Treating the prompt as a hypothesis means the output is evidence: you evaluate it (Discernment), diagnose what the prompt under- or mis-specified, and revise the description — that cycle is the Description–Discernment loop. (A) front-loads drafting without any feedback signal. (C) misattributes the loop to something Claude does internally rather than a human practice. (D) confuses the loop with a rule about prompt length.
Q3: A. Decomposition means splitting a big request into sequenced smaller ones — outline, then sections with review between steps, then synthesis — so each step gets focused attention and you can steer before errors compound. (D) is the runner-up: it splits the work but sacrifices shared context across competitors and pushes the synthesis burden back onto the human. (B) makes the single monolithic prompt bigger rather than smaller, and (C) mistakes effort for structure — no instruction to "take its time" restructures the task.
Q4: D. Steerability cuts both ways: vague input yields generic output, because Claude generates plausible text from whatever steering it has — here, almost none. (A) misapplies the knowledge cutoff, which limits knowledge of recent facts, not the ability to write a post from details you supply. (B) encodes the misconception that Claude browses the web by default; it doesn't. (C) encodes the "copy-paste from training data" misconception — the model predicts plausible continuations rather than retrieving a stored post.
Q5: B. Intermittently wrong format is the signature of a format described loosely or not at all; the fix is to specify the exact structure and show it — an example row beats a description of one. (D) is the runner-up but "structured output" is still ambiguous: bullets, headings, and lists are all "structured," so drift continues. (A) explains why consistency matters without defining what consistent looks like, and (C) delegates the very decision that needs pinning down.
Q6: C. Chaining wins when steps are dependent and intermediate review lets you catch and steer problems before they propagate — that's the point of sequencing. (A) and (B) are arbitrary proxies: length and source count don't by themselves create dependencies. (D) is false as stated — shorter prompts aren't inherently less error-prone; decomposition helps because of review points, not brevity.
Q7: D. With a huge document in context, instructions get diluted (steerability × working memory), and the guidance is to restate key constraints near the end where they won't get lost. (A) is the tempting runner-up: Claude will happily confirm it read the rules, but confirmation doesn't counteract dilution when it generates the summary. (B) buries the rules deeper in the noise, and (C) mistakes tone for placement — formality doesn't change where the rules sit in the context.
Q8: A. Iteration means using what the failed output reveals to revise the description; resampling the same prompt leaves the diagnosis unused, so improvement is luck rather than method. (B) is false — any model produces varied samples, and capability isn't the issue. (C) overstates a side effect: retries aren't primarily a context problem. (D) encodes the misconception that Claude learns from rejected drafts within a conversation and adapts its search — regeneration doesn't work that way.
Q9: B. Examples beat description for tone: two or three real on-voice posts show what adjectives can't pin down. (A) doubles down on the approach that's already failing. (C) is the runner-up — role assignment is a legitimate lever, but a role frames perspective without encoding this brand's specific voice, which examples capture directly. (D) assumes Claude will browse the site on its own; the reliable move is to put the material in the prompt yourself.
Q10: A. The canonical order starts with "be clear and direct" — state task, constraints, and expected format before reaching for anything else, because most failures are specification failures. Examples (B), chain-of-thought (C), and role assignment (D) are all real levers, but they come later in the list; applying them before clarifying the task decorates an ambiguous request instead of fixing it.
Q11: D. Cowork is delegate-and-review: the best request gives the task loop everything up front — goal, classification criteria, safety constraint (work on copies, per the start-with-reversible-tasks practice), an escalation path for edge cases, and a defined report. (A) is the runner-up: reviewing afterward is right, but "use your judgment" withholds the criteria and constraints the loop needs, inviting irreversible surprises. (B) collapses the delegation — the human does the classifying. (C) substitutes a generic norm for the team's actual intent.
Q12: C. Claude generated a plausible summary because it had nothing to ground on — the fix is to provide the source in context and constrain Claude to answer only from it. (D) is the runner-up: citations help surface unsupported claims, but with no document supplied there is nothing real to cite, and fabricated citations can look as confident as fabricated policy. (A) has Claude verify against itself rather than against a source, and (B) is wrong-layer — the context isn't stale; the information was never there.
Q13: A. This request needs an integration built — scheduled database access and automated email are beyond the chat surface — so the associate-correct move is to scope it, hand the technical piece to developers, and keep using Claude for what prompting genuinely covers: drafting the summary from sample data. (B) treats a missing capability as a wording problem; no iteration grants database access. (C) violates data-sensitivity basics. (D) is the runner-up misconception: Project instructions persist context across conversations, but they don't schedule or execute anything on their own.