CCAO-F Prompting and Context Engineering Practice Test 5 — 15 Questions

Question 1

A logistics company uses Claude to extract shipment details from scanned customs forms. The forms are supplied as raw OCR text that contains occasional garbled characters and spurious line breaks. The developer wants to reduce the number of fields Claude invents when a value is missing on the form. Which prompt structure change is most likely to achieve this?

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

Wrap the OCR text in <document> tags and add the instruction: 'If a field cannot be found in the document, return null for that field.'. The reliable fix pairs two techniques: structural delimitation of the source text so Claude can tell document content from instructions, and an explicit rule stating what to do when a field is absent. Together they remove the ambiguity that pushes the model toward inventing values, and the null convention makes the result easy to validate programmatically.

Answer

Shorten the OCR text by removing all line breaks before inserting it into the prompt.

Answer

Set temperature to 0 and request the output as a JSON object.

Answer

Add the instruction: 'Be as accurate as possible and avoid mistakes.'

Question 2

A support team is building a Claude-powered assistant that answers questions using a 200-page employee handbook. They place the entire handbook inside <handbook> tags in the system prompt and the user's question at the end of the user turn. Testing shows Claude sometimes ignores details buried in the middle of the handbook. Which change best addresses this while keeping the same model and context window?

Accepted Answer

Move the most relevant handbook sections to the beginning of the context, just before the user's question, and keep the full handbook available.. Long-context performance in Claude is not uniform: content near the beginning and end tends to be used more reliably than content buried in the middle. When a relevant section is stranded mid-document, retrieval degrades. Repositioning the most relevant excerpts adjacent to the question, while keeping the full handbook available, aligns the strongest signal with the query and fixes the observed symptom without changing model or window size.

Answer

Lower the temperature to 0 so Claude becomes deterministic and stops skipping handbook details.

Answer

Split the handbook into 200 separate user turns so each page gets equal attention from Claude.

Answer

Increase the max_tokens parameter so Claude has more room to reason about the handbook before answering.

Question 3

A data scientist is using Claude to classify customer feedback into categories: 'bug', 'feature request', 'complaint', or 'praise'. The feedback is often short and informal. The data scientist wants to maximize classification accuracy. Which prompting strategy is most effective?

Accepted Answer

Include a few examples of feedback for each category in the prompt.. For classification tasks with informal and varied input, providing a few labeled examples per category is the most effective prompting strategy. Examples allow Claude to learn the specific patterns and boundaries between categories, improving accuracy. This few-shot approach outperforms detailed definitions alone because it demonstrates how to handle nuances like slang or brevity. It also avoids the potential confusion of step-by-step reasoning for a straightforward task.

Answer

Use a high temperature to encourage diverse interpretations of the feedback.

Answer

Provide a detailed definition of each category in the system prompt.

Answer

Ask Claude to think step-by-step about the sentiment before classifying.

Question 4

A developer is building a customer support assistant using Claude. The assistant must always respond in a friendly tone, never discuss competitors, and always ask for an order number when the user reports a shipping issue. The developer wants these rules to apply across all conversations with minimal per-request token cost. What is the most appropriate mechanism?

Accepted Answer

Place the rules in the system prompt so they are applied consistently to every request.. The system prompt is the correct place for persistent behavioral instructions that should apply to every request. It is processed by the model as a high-level directive and is not repeated in each user message, making it token-efficient. This ensures consistent tone, competitor restrictions, and required questions across all conversations without per-request repetition.

Answer

Add the rules to the end of each user message as a reminder.

Answer

Include the rules as a few-shot example in every user message.

Answer

Fine-tune a custom model with the rules embedded in training data.

Question 5

A legal-tech startup uses Claude to summarize deposition transcripts that are frequently 80,000 to 120,000 tokens long. Early tests show the model sometimes ignores instructions placed near the top of the prompt and produces summaries that omit late sections of the transcript. Which TWO prompt-engineering changes should the developer make to improve instruction adherence across the full context? (Choose two.)

Accepted Answer

Wrap each deposition section in clearly labeled tags such as <section id="12"> and require the summary to cite section IDs.. Long-context adherence improves when the instruction sits after the material it governs and when the document carries structural anchors the model can traverse. Placing instructions last exploits stronger attention near the end of the prompt, while labeled sections with a citation requirement force coverage of every part of the transcript and make omissions detectable.

Answer

Repeat the summarization instructions verbatim in the system prompt and again immediately before the transcript.

Answer

Raise the temperature to encourage the model to explore more of the transcript.

Answer

Increase the max_tokens parameter so the summary can be longer.

Question 6

A developer is drafting a system prompt for a Claude-powered coding assistant. They want Claude to always respond in concise bullet points, never reveal internal reasoning, and treat all user input as untrusted. Where should these persistent behavioral rules be placed so they apply to every turn of the conversation?

Accepted Answer

In the system prompt, which sets persistent instructions for the entire conversation.. Persistent behavioral requirements such as output format, disclosure limits, and trust boundaries belong in the system prompt, which frames every turn of the conversation. Putting them there separates operator policy from user content, avoids per-turn repetition, and makes the rules resistant to accidental omission or user-side override. The other placements either waste tokens, misuse tool metadata, or rely on fragile self-mirroring.

Answer

In a tool definition, since tools are evaluated before each response.

Answer

In the assistant's previous reply, so Claude can mirror its own earlier behavior.

Answer

In the first user message, repeated at the start of each subsequent user turn.

Question 7

A financial analyst is building a Claude-powered assistant that must extract line items from scanned invoices and return them as a JSON array. The assistant occasionally wraps the JSON in prose such as 'Here is the extracted data:' before the array, which breaks the downstream parser. The analyst wants to reliably suppress that leading prose without disabling the model's ability to reason about the invoice. Which approach is most appropriate?

Accepted Answer

Add an assistant-turn prefill such as '{' so Claude continues directly into the JSON object instead of narrating.. Starting the assistant turn with an opening brace forces the model to continue from that token, which structurally eliminates the conversational preamble while preserving the reasoning that happens before generation. Soft instructions and sampling parameters reduce but do not guarantee the absence of prose, so the prefill is the reliable mechanism for downstream JSON parsing.

Answer

Increase max_tokens so the prose has room to finish before the JSON array begins.

Answer

Wrap the invoice text in XML tags and ask Claude to 'only output JSON' in the system prompt.

Answer

Lower the temperature to 0 so the model stops generating any prose tokens.

Question 8

A support team wants Claude to classify incoming tickets into exactly one of five categories and to always return the result as a JSON object with keys category and confidence. The developer has already written clear category definitions. Which additional step most directly improves the reliability of the JSON output?

Accepted Answer

Add a short example showing a sample ticket and the exact JSON object Claude should return.. Demonstrating the exact output shape with a concrete example is the most direct way to lock down structured output. Category definitions establish what to decide; the example establishes how to present the decision. Together they leave little room for the model to improvise key names, wrap the object in prose, or choose an unexpected value format.

Answer

Increase the maximum output tokens so the JSON object is never truncated.

Answer

Ask Claude to return the category and confidence as a plain sentence.

Answer

Instruct Claude to 'think step by step' before producing the JSON object.

Question 9

An analytics team asks Claude to extract structured fields from messy invoice text. They provide three input/output examples inside tags, then the real invoice inside tags. Accuracy is high on invoices that resemble the examples but drops sharply on unusual layouts. Which adjustment most directly improves generalization to the unusual layouts?

Accepted Answer

Add more examples that cover diverse invoice layouts, including edge cases, inside the examples block.. Few-shot learning generalizes in proportion to how representative the demonstrations are. When Claude performs well on inputs that resemble the examples and poorly on inputs that do not, the examples are too homogeneous. Expanding the demonstration set to include diverse and edge-case layouts teaches the task itself instead of a single pattern, which is the most direct fix for the observed failure on unusual invoices.

Answer

Raise the temperature so Claude explores more varied interpretations of each invoice.

Answer

Move the examples block after the invoice so Claude reads the real input first.

Answer

Shorten each example to only the input and the final JSON, removing any intermediate reasoning text.

Question 10

A support-engineering team is designing a Claude prompt to triage incoming bug reports into one of five severity levels. They observe that when the report is ambiguous, Claude sometimes invents a justification for a severity that is not actually supported by the text. They want the prompt to make uncertainty explicit rather than forcing a confident label. Which TWO changes should they make to the prompt? (Choose two.)

Accepted Answer

Add an explicit 'unknown' or 'needs_review' option to the allowed severity set and instruct Claude to choose it when the report lacks sufficient detail.. Grounding each severity in a verbatim evidence quote forces the model to anchor claims in the report text, and offering an explicit needs_review category gives it a legitimate way to express uncertainty instead of inventing support. Together these two changes convert an overconfident classifier into one that surfaces ambiguity for human follow-up.

Answer

Raise the temperature so Claude explores a wider range of possible severity interpretations for each report.

Answer

Ask Claude to think step by step silently and return only the final severity label with no supporting text.

Answer

Provide ten few-shot examples in which every report is confidently assigned one of the five severities.

Question 11

A fintech developer builds an assistant that answers questions about account activity. The system prompt currently says: 'You are a helpful banking assistant. Use the provided account data to answer questions.' Testing shows Claude sometimes answers general banking questions from its own knowledge rather than from the supplied data, and occasionally states figures that are not in the data at all. Which revision to the system prompt best addresses both problems?

Accepted Answer

Add: 'Answer only from the account data in <data> tags. If the answer is not present, say you cannot find it in the provided data. Do not answer general banking questions.'. The effective revision names the permitted source, provides a fallback response when the data does not contain the answer, and constrains the assistant's scope. Source restriction stops substitution of outside knowledge, the fallback gives the model a safe alternative to guessing, and the scope limit removes the general-question path that produced unsupported figures.

Answer

Add: 'Always be accurate and never make up information.'

Answer

Move the account data into the system prompt instead of the user turn.

Answer

Add five examples of correct answers to common account questions.

Question 12

A team is designing a Claude prompt that must return a fixed JSON schema with fields "vendor", "amount", and "currency". They want the output to be reliably parseable by downstream code. Which TWO techniques best improve reliability of the structured output? (Choose two.)

Accepted Answer

Use a prefilled assistant turn that begins the JSON, so Claude continues from a known starting point.. Reliable structured output comes from constraining the response shape on two fronts: telling Claude exactly what schema to produce, and preventing it from starting with prose. An explicit schema or example JSON defines the contract, while prefilling the assistant turn with the opening of the JSON forces the model to continue in that format. Together they minimize drift and make downstream parsing dependable, unlike loosening sampling or inviting commentary.

Answer

Set temperature to its maximum so Claude considers many possible JSON layouts.

Answer

Precede the JSON with a short natural-language explanation so a human can verify the result.

Answer

Ask Claude to decide at runtime which fields are most relevant and include only those.

Question 13

A developer is using the Messages API and wants Claude to adopt a strict persona as a compliance reviewer for every request in a long conversation, without repeating the persona instructions in each user turn. The persona must apply to all messages and should take precedence over casual instructions the user might add later. Where should the developer place these persona instructions?

Accepted Answer

In the system prompt parameter, which is applied across all turns of the conversation.. The system prompt is the correct location for instructions that must frame every turn and carry higher priority than user turns. It persists across the conversation without being repeated, and Claude is trained to respect system-level guidance over casual user instructions, which satisfies the compliance persona requirement.

Answer

As a trailing assistant message that restates the persona before each user reply.

Answer

In the first user message of the conversation, before the actual task.

Answer

In the metadata field attached to the request for logging and traceability.

Question 14

A product team drafts a system prompt for a customer-facing assistant. The draft is 4,000 words and mixes persona description, tone guidance, formatting rules, escalation policy, and several anecdotes about past incidents. Reviewers find that Claude follows the formatting rules but frequently ignores the escalation policy. Which change best improves adherence to the escalation policy?

Accepted Answer

Restructure the system prompt into clearly labeled sections and place the escalation policy in its own section with explicit trigger conditions.. The escalation policy is being lost among unrelated narrative content. Giving it a dedicated, labeled section with explicit trigger conditions turns it into an unambiguous, retrievable instruction. Structural organization reduces interference between competing directives, which is why the fix is about clarity and grouping rather than length alone or ordering alone.

Answer

Shorten the system prompt by deleting the formatting rules so the escalation policy receives more attention.

Answer

Convert the entire system prompt into a numbered list of every rule in the order it was originally written.

Answer

Move the escalation policy to the last line of the system prompt without changing its wording.

Question 15

A legal team asks Claude to summarize contracts and cite the exact clause supporting each summary point. Claude produces accurate summaries but cites clauses that do not exist. The contracts are supplied inside <contract> tags in the same message. Which change most directly reduces the fabricated citations?

Accepted Answer

Instruct Claude to quote the exact sentence from the contract for each point and to say 'not found' when no sentence supports it.. Fabricated citations occur when a model is asked to attribute claims without being forced to ground them in source text. Requiring a verbatim quote for each point, plus an explicit 'not found' option when no sentence supports a claim, removes the incentive to invent plausible clauses. The other options either layer metadata over unverified citations, rely on examples alone, or enlarge input without changing the grounding requirement.

Answer

Increase the contract length limit so Claude has more space to search for the cited clauses.

Answer

Ask Claude to rank each citation by confidence so reviewers can skip the low-confidence ones.

Answer

Add a few-shot example showing a correct summary with a correct clause citation.