20 Questions — Domain 1 : Prompting and Task Execution

Expand any question to reveal the correct answer and explanation.

1. 1 A marketing team needs to transform a 50-page technical white paper into a 10-slide executive presentation. According to CCAO-F Domain 1 principles, what is the most effective way to structure this prompt sequence to ensure high accuracy?

   Consider the benefits of breaking a large, multi-stage project into sequential operations.

   Decompose the task by first prompting Claude to extract key themes and data points before generating the slide content.

   Decomposition ensures the model focuses on information extraction before shifting to the creative synthesis phase, improving fidelity.

   • ✗ Paste the entire white paper into a single prompt along with the slide requirements and speaker note constraints.

     Attempting a single-pass synthesis for complex tasks often results in context overload and omitted details.

   • ✗ Ask Claude to summarize the paper in one paragraph, then use that paragraph as the only source for the 10 slides.

     Relying on a single-paragraph summary for a 10-slide deck leads to significant loss of information density and nuance.

   • ✗ Provide the paper and ask Claude to generate the presentation in Research Mode without specific slide instructions.

     Research Mode is for gathering external info; it does not replace the need for structured instructions and task decomposition.

2. 2 During a multi-hour iterative session, you notice Claude is beginning to lose track of earlier constraints and formatting decisions. What is the recommended strategy for managing this context decay?

   Think about how to retain core progress while removing the cumulative weight of a long conversation.

   Ask Claude to summarize the key decisions and context, then initiate a fresh chat session using that summary.

   This clears unnecessary 'noise' from the history while preserving the essential parameters needed to continue the work accurately.

   • ✗ Prompt Claude to 'remember all previous instructions' at the start of every new message.

     Repeatedly asking the model to remember does not clear the actual context window limit or reduce noise.

   • ✗ Continue the conversation but switch to the most capable model (Opus) to handle the increased history.

     Switching models does not solve context decay; the conversation history still consumes the context window regardless of the tier.

   • ✗ Manually delete the oldest messages in the chat history to free up tokens.

     Claude's standard interfaces generally do not support selective deletion of historical turns to manage the context window.

3. 3 When designing a few-shot prompt to help Claude classify ambiguous business communications, what is the most effective approach for selecting examples?

   Focus on the specific quantity and nature of examples that best improve model judgment.

   Provide 2-4 targeted examples that include ambiguous and edge cases along with the reasoning for the classification.

   Focused examples with reasoning help the model internalize the underlying logic for difficult or 'gray area' decisions.

   • ✗ Provide 10-15 simple, straightforward examples to maximize the volume of training data Claude receives.

     High volume is less effective than quality; 10-15 simple cases do not help the model resolve complex edge cases.

   • ✗ Only provide examples of the most common classification type to reinforce the standard output format.

     Focusing only on common cases prevents the model from learning how to differentiate between categories or handle rare inputs.

   • ✗ Include a single, perfect example of every possible category to ensure a broad but shallow reference set.

     Single examples often fail to capture the variability and nuance required for high-accuracy classification in realistic scenarios.

4. 4 A business practitioner wants to refine a prompt that is producing vague, generic reports. Which modification aligns with Domain 1's 'structured prompting' best practices?

   Identify the elements that provide objective boundaries and goals for the model's output.

   Specifying the target audience, the specific purpose, key metrics to include, and the required length.

   Structured prompts require explicit constraints and success criteria to move beyond generic outputs.

   • ✗ Adding a polite request for Claude to 'do its best' and 'be as detailed as possible'.

     Politeness and vague adjectives like 'detailed' do not provide the concrete parameters the model needs for accuracy.

   • ✗ Switching the prompt to a new chat session to ensure no previous context interferes.

     While a fresh chat avoids context noise, it does not fix a prompt that is fundamentally lacking in structure and parameters.

   • ✗ Increasing the temperature setting in the Claude Console to encourage more creative report generation.

     CCAO-F focuses on business productivity where accuracy and structure are prioritized over the randomness associated with higher temperature.

5. 5 Which task adaptation requires moving from 'drafting' instructions to 'objective analysis' instructions?

   Consider what structural change forces a shift from 'how it is said' to 'how it is judged'.

   Shifting focus from sentence variety and brand tone to explicit evidence-based evaluation criteria.

   Objective analysis necessitates a framework for judgment, whereas drafting focuses on presentation and style.

   • ✗ Reducing the number of constraints to allow for more open-ended exploration of the topic.

     Objective analysis typically requires more rigid constraints to ensure the findings remain grounded in facts.

   • ✗ Increasing the use of XML tags to better structure the final document's headings and subheadings.

     XML tags are a formatting tool; they do not inherently change the cognitive nature of the task from drafting to analysis.

   • ✗ Providing more few-shot examples of creative writing styles to influence the model's output.

     Examples of creative styles are useful for drafting but do not assist with the evaluative logic of objective analysis.

6. 6 You are iterating on a prompt to resolve a consistent error where Claude includes technical jargon in emails meant for a non-technical audience. What is the best corrective action?

   Look for a professional method of defining constraints that doesn't rely on colloquialisms or repetition.

   Provide explicit definitions of the audience's literacy level and a list of specific jargon terms to avoid.

   Defining the audience and providing negative constraints (terms to avoid) provides clear, professional boundaries.

   • ✗ Prompt Claude to 'explain like I'm five' (ELI5) to ensure maximum simplicity.

     Simplified tropes like ELI5 can result in patronizing or overly simplistic output that is inappropriate for a business context.

   • ✗ Ask Claude to self-evaluate its jargon use on a scale of 1 to 10 and only output if the score is below 3.

     Self-evaluation scores are not reliable indicators of accuracy or adherence to constraints.

   • ✗ Repeat the instruction 'Do not use jargon' three times at the end of the prompt for emphasis.

     Simple repetition is less effective than providing concrete definitions and examples of the desired constraint.

7. 7 An Associate is tasked with reconciliation of two conflicting Q3 data reports. Why does the CCAO-F exam guide suggest that a single prompt may be insufficient for this task?

   Consider the cognitive load placed on the model when asked to synthesize and generate at the same time.

   Synthesis and reconciliation are highly complex and prone to errors if performed simultaneously with drafting.

   Decomposing the task into 'identify conflicts' then 'propose reconciliation' then 'draft summary' ensures higher fidelity at each stage.

   • ✗ Claude cannot process two separate files in a single prompt effectively.

     Claude is capable of processing multiple files; the issue is the cognitive complexity of the task, not the file count.

   • ✗ The reconciliation requires a different model (Haiku) than the drafting task (Opus).

     While model tiering is a factor, the fundamental issue is the logical sequence of the operation.

   • ✗ Multi-file reconciliation is a technical task that should always be escalated to a Developer.

     Reconciliation is a core knowledge task suitable for an Associate; only the programmatic integration would require escalation.

8. 8 When configuring a Claude Project for a recurring weekly campaign recap, what is the best practice for maintaining instructions and knowledge sources?

   Focus on the role of human oversight in maintaining a persistent, reliable workspace.

   Establish a dedicated project owner to update the standing instructions and remove outdated files on a set schedule.

   Active maintenance ensures the project remains a 'source of truth' with current instructions and relevant knowledge.

   • ✗ Paste the latest week's data and new instructions into a fresh chat within the project every Monday.

     This manual approach negates the benefits of persistent configuration and increases the risk of inconsistent prompting.

   • ✗ Use a single, long-running chat thread to ensure Claude has access to every previous week's performance data.

     This leads to context decay and noise, as older performance data may become irrelevant and clutter the context window.

   • ✗ Instruct Claude to automatically delete any files in the knowledge base that it identifies as 'outdated'.

     Claude cannot autonomously manage the file repository of its own project workspace; this is a human configuration task.

9. 9 In a brainstorming prompt designed to generate new product ideas, which structural element is most appropriate compared to a drafting prompt?

   Think about whether you want the model to be more expansive or more restrictive for this specific task.

   Prioritizing broad exploration and quantity of responses over specific formatting and evidence synthesis.

   Brainstorming prompts should encourage divergent thinking, whereas drafting prompts are about convergent, structured output.

   • ✗ Strict adherence to a specific JSON output schema to ensure data compatibility.

     Rigid schemas in brainstorming can stifle the model's creative output and limit the range of ideas generated.

   • ✗ Requiring a citation for every idea generated to ensure feasibility.

     Requiring citations for new ideas is illogical and forces the model toward hallucination or repetitive, known concepts.

   • ✗ Decomposing the task into five sequential stages to ensure each idea is fully vetted before the next is generated.

     Stage-based vetting is for analysis/development; initial brainstorming benefits from lower initial friction.

10. 10 A user wants Claude to write an email based on a sensitive meeting transcript but must follow organizational privacy policies. What is the most appropriate 'Task Execution' step?

    Identify the safest stage in the workflow to address data sensitivity concerns.

    Anonymize or redact the transcript manually to remove regulated data before including it in the prompt.

    Handling data sensitivity at the input stage is the most reliable way to ensure compliance with privacy standards.

    • ✗ Ask Claude to 'redact all sensitive information' while it drafts the email.

      Relying on Claude to identify and redact sensitive info in-stream is risky; the data has already been shared with the model.

    • ✗ Include a policy document in the Claude Project and tell Claude to 'comply with all rules'.

      General instructions to 'comply' are often ignored or misinterpreted without specific, hands-on data handling.

    • ✗ Password-protect the document before uploading it to the Claude Project knowledge base.

      Claude generally cannot access password-protected files, and this does not address the core data-handling requirement.

11. 11 You are creating a prompt for a high-scrutiny task that requires a specific professional tone. Which addition is most likely to produce the desired result consistently?

    Think about the difference between telling the model what you want versus showing it.

    2-4 few-shot examples that demonstrate both the target tone and the reasoning behind stylistic choices.

    Examples provide the most effective stylistic signal, especially when they include the logic for specific choices.

    • ✗ A list of 50 adjectives describing the tone (e.g., 'earnest', 'stately', 'direct').

      Overloading with adjectives creates conflicting signals and does not provide concrete examples for the model to follow.

    • ✗ A system instruction to 'assume the persona of a world-class CEO'.

      Personas are often generic and can introduce unwanted tropes rather than specific professional tone constraints.

    • ✗ An instruction to 'use the same tone as the provided source documents' without examples.

      Source documents may contain multiple tones; without examples, the model may anchor to the wrong stylistic elements.

12. 12 Which scenario described below illustrates the 'vague request' anti-pattern discussed in Domain 1 of the CCAO-F exam?

    Look for the request that leaves the most room for the model to guess the user's intent and parameters.

    'Take this project plan and write something about our upcoming goals for the next meeting.'

    This is vague because 'write something' provides no constraints on audience, specific focus, length, or desired tone.

    • ✗ 'Write a summary of the Q3 regional sales report for the operations team, focusing on supply chain delays.'

      This request includes a specific source (Q3 report), target audience (operations), and clear focus area.

    • ✗ 'Analyze these three transcripts and list every mention of pricing in a table with two columns: Vendor and Proposed Price.'

      This request has clear source material, a specific task (listing pricing), and a defined output format (table).

    • ✗ 'Draft a LinkedIn post announcing our new partnership, under 200 words, using an enthusiastic but professional tone.'

      This request defines the platform (LinkedIn), task (announcement), length constraint, and tone.

13. 13 A team finds that Claude's output is consistently incomplete when tasked with summarizing long research papers. What is the best task-decomposition fix?

    Consider how to ensure every part of a long source document receives equal attention.

    Instruct Claude to 'summarize each section individually' before synthesizing the final summary.

    Sectional decomposition prevents the model from glossing over later parts of the document due to context processing limits.

    • ✗ Increase the output length constraint to 'at least 5,000 words' in the prompt.

      Arbitrary length constraints do not guarantee information completeness and can lead to filler or repetition.

    • ✗ Switch to a faster model (Haiku) to allow the summary to be generated more quickly.

      Speed is irrelevant to completeness; actually, a faster model might be less thorough for complex synthesis.

    • ✗ Provide 10 examples of good summaries to guide the model's style.

      Style examples help with tone, but they do not solve the structural issue of information loss in long-document processing.

14. 14 When iterating on a prompt based on an unsatisfactory response, why is it recommended to maintain a 'prompt log'?

    Think about how documenting your changes helps you be more systematic in your testing.

    To identify which specific adjustments improve results and avoid repeating failed edits.

    A log prevents circular trial-and-error and allows the practitioner to see patterns in how Claude responds to certain constraints.

    • ✗ To provide a history for the technical team when you escalate the task to a Developer.

      While useful for escalation, the primary benefit is for the Associate's own iterative process.

    • ✗ Claude can read the prompt log to learn from your previous mistakes and automatically improve.

      Claude does not have cross-session memory of your local files or logs unless you explicitly provide them in every chat.

    • ✗ The prompt log is a mandatory requirement for the CCAO-F certification renewal assessment.

      Renewal is via a non-proctored assessment; maintaining a personal log is a best practice, not a formal program requirement.

15. 15 An Associate is using Claude for research. To balance quality and speed, which model tiering strategy should they use according to the 'Cost-Quality Optimization' principle?

    Determine which approach allows for professional quality while being mindful of operational costs.

    Use a lighter model (Haiku) for high-volume data sorting, and reserve the premium model (Opus) for final synthesis.

    Matching the model to the task's complexity optimizes both budget and processing speed without sacrificing final output quality.

    • ✗ Use the most expensive model (Opus) for every step to ensure the highest possible quality for the research.

      Always using the top tier is financially inefficient and often unnecessary for simple, preliminary steps.

    • ✗ Standardize the entire workflow on the fastest model (Haiku) to maximize efficiency.

      Relying solely on the fastest model may lead to lower quality in the complex reasoning required for final synthesis.

    • ✗ Allow Claude to autonomously decide which model to use for each step of the research.

      Claude does not have the capability to switch its own underlying model tier mid-session based on task difficulty.

16. 16 What is the primary distinction between 'Prompt-based' enforcement and 'Programmatic' enforcement in an AI workflow?

    Consider the level of risk you are willing to accept if an instruction is ignored.

    Prompt-based enforcement is used for stylistic choices, while programmatic enforcement is used for financial or safety-critical steps.

    Prompts are probabilistic (not 100% reliable), making them unsuitable for rules with significant legal or safety consequences.

    • ✗ Programmatic enforcement is only available to Developers, while Associates must rely solely on prompts.

      An Associate advises on workflows and must know when to recommend programmatic gates, even if they don't code them.

    • ✗ Prompt-based enforcement is faster to implement and therefore always preferred in an agile business environment.

      Speed of implementation does not outweigh the need for deterministic reliability in high-stakes operations.

    • ✗ There is no functional difference; they are two different terms for the same instructional technique.

      The difference is fundamental: one relies on model interpretation, the other on code-based hard limits (hooks/gates).

17. 17 A user needs Claude to extract data from multiple inconsistent source formats (PDFs, spreadsheets, and emails). Which prompting technique is most impactful here?

    Identify the technique that helps the model recognize patterns across different input structures.

    Providing 2-4 few-shot examples showing the extraction from each of the varied source structures.

    Few-shot examples are highly effective at teaching the model how to map inconsistent inputs to a consistent output format.

    • ✗ Asking Claude to 'be extra careful with the different formats'.

      Abstract warnings do not provide the model with the structural guidance needed for varied data extraction.

    • ✗ Telling Claude to 'convert everything to a PDF first' before extracting.

      Claude cannot autonomously perform file format conversions; it can only interpret the content of the files as they are provided.

    • ✗ Using a separate chat session for each file type to avoid confusing the model.

      This adds massive manual overhead and prevents the model from producing a single, consolidated output from all sources.

18. 18 In the context of the CCAO-F exam, what is a 'defining Associate skill' regarding task execution and role boundaries?

    Think about how an Associate interacts with more technical roles in an organization.

    Recognizing the edge of the role and knowing when to escalate complex implementations to a Developer or Architect.

    Associates must know when a task requires technical expertise (like API code or agent design) beyond their productivity-focused scope.

    • ✗ The ability to write basic Python scripts to automate prompt delivery.

      API and software development skills are specifically excluded from the Associate role's requirements.

    • ✗ Memorizing every available Claude feature and capability for a comprehensive feature-list recall.

      The exam tests practical judgment in workplace scenarios, not rote memorization of feature lists.

    • ✗ Ensuring that every AI-generated output is used immediately to maximize organizational efficiency.

      Associates are explicitly taught to apply fact-checking and determine when human review is required before use.

19. 19 An Associate is redesigning a workflow where a team manually reformats data across multiple documents. According to the 'Process Redesign' principle, what should be their primary goal?

    Determine whether you should focus on replicating a legacy process or rethinking it.

    Look for opportunities to redesign the entire sequence around Claude's strengths, potentially eliminating sequential turns.

    True optimization often involves restructuring the workflow to leverage AI for synthesis rather than just simple automation of manual tasks.

    • ✗ Automate the existing manual steps exactly as they currently exist to ensure no disruption.

      Automating inefficient legacy steps misses the opportunity to simplify the entire process using AI capabilities.

    • ✗ Ensure that Claude performs every single step of the process with zero human intervention.

      Redesign must still include appropriate review and verification checkpoints to manage AI limitations.

    • ✗ Add more steps to the workflow to provide Claude with more granular instructions for every minute task.

      Excessive granularity can lead to 'instruction bloat' and context noise, making the process less efficient.

20. 20 When generating a complex research report, an Associate needs to ensure Claude doesn't present fabricated citations. What is the most effective way to address this in the prompting phase?

    Focus on how to provide the model with the 'ground truth' it should use for the task.

    Provide the specific source documents in the prompt and instruct Claude to 'only use information from these files'.

    Grounding the model in specific, provided knowledge sources is the first and most critical step in preventing hallucinations.

    • ✗ Ask Claude to provide a 'confidence score' for each citation it includes.

      Self-reported confidence is not a reliable indicator of whether a detail like a citation was hallucinated.

    • ✗ Tell Claude that 'hallucinating is strictly prohibited by organizational policy'.

      The model cannot comply with abstract 'prohibitions' against its own probabilistic errors; it needs structural grounding.

    • ✗ Use the fastest model (Haiku) to minimize the chance of 'overthinking' and inventing details.

      Lighter models are generally *more* prone to hallucinations in complex research tasks than the premium tier.