Skip to content
Cheat sheets

CCAR-F · Domain 4: Prompt Engineering & Structured Output · 20% of the exam

Task 4.2: Apply few-shot prompting to improve output consistency and quality

A few targeted examples that show the format, the reasoning behind ambiguous calls, acceptable versus problematic patterns, and how to extract from documents that are laid out differently.

Study it

Sample questions

Try each one before opening the answer. Every option is explained, with the Anthropic documentation page that proves it.

Question 1 · choose 1

Your CI reviewer writes a markdown summary for each pull request. The instructions ask for the location, the issue, a severity and a suggested fix for every finding, yet some findings lack line numbers, severities vary in wording, and fixes are sometimes missing. Which change most reliably makes the findings consistent?

  1. AAdd a few example findings to the prompt, each written in the exact format
  2. BSet the temperature to 0 so that every finding comes out in the same format each time the review runs
  3. CRepeat the format requirements in capital letters at both the start and the end of the instructions
  4. DAsk Claude to check its findings against the format rules before finishing
Show the answer and why

  • AAdd a few example findings to the prompt, each written in the exact format

    Correct

    Examples are one of the most reliable ways to steer output format; they show the structure the instructions only describe.

  • BSet the temperature to 0 so that every finding comes out in the same format each time the review runs

    Incorrect

    Lower randomness does not teach the format; it can make the same inconsistent layout repeat more reliably.

  • CRepeat the format requirements in capital letters at both the start and the end of the instructions

    Incorrect

    Repetition and emphasis do not show what a correct finding looks like, so the variation continues.

  • DAsk Claude to check its findings against the format rules before finishing

    Incorrect

    A self-check against rules it already misapplies is weaker than a concrete pattern to copy.

When detailed instructions alone give inconsistent formatting, add a few targeted examples of the exact output. Few-shot examples demonstrate the format in a way prose cannot.

Anthropic documentation

Task 4.2Report a problem with this question

Question 2 · choose 1

Your extraction system pulls cited references out of research reports. Some reports cite inline ("Smith et al., 2021"), others use numbered markers that point to a bibliography at the end. For the second kind, the references field often comes back empty although the bibliography is in the document. What should you change first?

  1. ARetry every extraction whose references field comes back empty, with the same prompt and document
  2. BMake the references field required so Claude must fill it for each report
  3. CShorten the documents by removing their bibliographies before extraction so the prompt focuses on the body
  4. DAdd few-shot examples that show correct extraction from both citation layouts
Show the answer and why

  • ARetry every extraction whose references field comes back empty, with the same prompt and document

    Incorrect

    An identical retry repeats the same misreading of the layout.

  • BMake the references field required so Claude must fill it for each report

    Incorrect

    Requiring a value pushes Claude to produce something, including invented references, without showing it where to look.

  • CShorten the documents by removing their bibliographies before extraction so the prompt focuses on the body

    Incorrect

    This deletes the very section that holds the references for the numbered-marker reports.

  • DAdd few-shot examples that show correct extraction from both citation layouts

    Correct

    Examples of each document structure show Claude how to find the data in layouts it is currently missing.

Few-shot examples that cover varied document structures (inline citations versus bibliographies) reduce empty or null extractions of fields that are present but laid out differently.

Anthropic documentation

Task 4.2Report a problem with this question

Question 3 · choose 1

Your extraction prompt for shipping manifests already has detailed instructions, and output is still inconsistent. You are adding few-shot examples for the first time and want good results without inflating every request. How many examples does Anthropic's prompting guidance suggest as a starting point?

  1. AThree to five well-chosen, diverse examples
  2. BOne example per field in the schema, about 25 in all
  3. CA single example of the most common manifest layout
  4. DNone at first; keep adding instruction rules until the output stabilizes
Show the answer and why

  • AThree to five well-chosen, diverse examples

    Correct

    A few relevant, diverse examples give the best results for the tokens they cost.

  • BOne example per field in the schema, about 25 in all

    Incorrect

    That many examples inflates every request without proportional benefit.

  • CA single example of the most common manifest layout

    Incorrect

    One example is easily over-copied and shows none of the variation in real manifests.

  • DNone at first; keep adding instruction rules until the output stabilizes

    Incorrect

    More rules is what has not worked; examples are one of the most reliable ways to steer format and structure.

Examples are one of the most reliable ways to steer output. The prompting guidance suggests three to five for best results, rather than one example or dozens.

Anthropic documentation

Task 4.2Report a problem with this question

Question 4 · choose 1

You have three examples for your extraction prompt and suspect they are too similar. You would like more coverage without writing every example by hand. What does the guidance suggest?

  1. ARemove the examples and add more rules instead
  2. BCopy each example three times with small wording changes
  3. CAsk Claude to check them for diversity and draft more
  4. DUse only examples taken from a different task
Show the answer and why

  • ARemove the examples and add more rules instead

    Incorrect

    Rules do not replace the demonstration examples give.

  • BCopy each example three times with small wording changes

    Incorrect

    Near-duplicates reinforce the same narrow pattern.

  • CAsk Claude to check them for diversity and draft more

    Correct

    Claude can evaluate examples for relevance and diversity or generate additional ones from your set.

  • DUse only examples taken from a different task

    Incorrect

    Examples should mirror your own use case.

You can ask Claude to evaluate your examples for relevance and diversity, or to generate additional ones based on your initial set.

Anthropic documentation

Task 4.2Report a problem with this question

Question 5 · choose 1

Your support prompt has grown a list of 70 edge-case rules ("if the customer mentions X, do Y"). Behavior is brittle, and new edge cases keep breaking it. What does Anthropic's context-engineering guidance recommend instead?

  1. AA small set of diverse, canonical examples
  2. BOne very long example that covers all 70 cases at once
  3. CRemoving all guidance and letting the model decide
  4. DMore rules, so that every new edge case is covered
Show the answer and why

  • AA small set of diverse, canonical examples

    Correct

    Canonical examples portray the behavior and generalize better than a laundry list of rules.

  • BOne very long example that covers all 70 cases at once

    Incorrect

    A single sprawling example is hard to learn from.

  • CRemoving all guidance and letting the model decide

    Incorrect

    Too little guidance fails in a different way.

  • DMore rules, so that every new edge case is covered

    Incorrect

    The list keeps growing and still breaks on the next case.

Prefer a curated set of diverse, canonical examples over stuffing a prompt with edge-case rules.

Anthropic documentation

Task 4.2Report a problem with this question

Question 6 · choose 1

Your few-shot examples for extracting fields from shipping manifests use tidy, typed documents, while real manifests are scanned, abbreviated and full of handwritten notes. Results on real documents are poor. What should you change?

  1. AAdd a note that real documents will look different
  2. BUse examples that look like the real manifests
  3. CSwitch the examples to invoices, which are better structured
  4. DKeep the tidy examples, since they show the ideal output most clearly
Show the answer and why

  • AAdd a note that real documents will look different

    Incorrect

    A note does not show how to handle the differences.

  • BUse examples that look like the real manifests

    Correct

    Examples should closely match the actual inputs to be useful.

  • CSwitch the examples to invoices, which are better structured

    Incorrect

    Examples from another document type are less relevant.

  • DKeep the tidy examples, since they show the ideal output most clearly

    Incorrect

    The output can be ideal while the inputs look realistic.

Make examples relevant: they should mirror the real inputs, including their mess, as well as cover diverse cases.

Anthropic documentation

Task 4.2Report a problem with this question

Question 7 · choose 1

Your support replies are accurate but swing between stiff and overly casual. The brand guide describes the tone in a paragraph. What most reliably makes the tone consistent?

  1. AA longer paragraph describing the brand voice
  2. BA list of words the agent must never use
  3. CA temperature of 0 for every reply
  4. DA few example replies in the target tone
Show the answer and why

  • AA longer paragraph describing the brand voice

    Incorrect

    Description alone is what produces the inconsistency now.

  • BA list of words the agent must never use

    Incorrect

    Prohibitions do not show what the tone should be.

  • CA temperature of 0 for every reply

    Incorrect

    Lower randomness does not define the target tone.

  • DA few example replies in the target tone

    Correct

    Examples steer tone and style more reliably than a description.

Few-shot examples steer format, tone and structure; wrap them in example tags so Claude can tell them apart from instructions.

Anthropic documentation

Task 4.2Report a problem with this question

Practise domain 4 →Practise all domains →