CCAR-P - Claude Models, Prompting & Context Engineering (13% of the exam) - Section 2.3

Apply prompt engineering techniques (zero-shot, few-shot, chain-of-thought).

Choosing between zero-shot instructions, few-shot examples and step-by-step reasoning for a given task. Candidates should know that examples steer format and edge-case handling, that reasoning helps multi-step problems at a cost in tokens and latency, and that a clear instruction alone often suffices.

zero-shotfew-shot exampleschain-of-thoughttoken and latency cost of reasoning

Practice question for this objective

Free sampleClaude Models, Prompting & Context Engineeringeasy

A software company's incident tool uses Claude to turn on-call engineers' notes into public status-page updates. Its prompt holds a written instruction and four worked examples, which were added a year ago because zero-shot updates varied widely in tone and length. Last month the communications team changed the house format from a single paragraph to three labelled lines (impact, current status, time of next update) and rewrote the instruction to match, but left the four examples unchanged. Of 300 updates since, 71 per cent still arrive as a single paragraph. The fix must ship this week, and the platform team will not add model calls per update or change the model tier. What should the architect recommend?

ARewrite the four worked examples in the three-line format so they match the new instruction Correct
BRestate the three-line rule in capitals at the end of the prompt so it outweighs the examples
CRemove the four worked examples and let the rewritten instruction set the format by itself
DAdd a second call that checks each update and regenerates any not written in three lines

When few-shot examples contradict a changed instruction, the examples usually win on format, so update the examples rather than shouting the instruction or adding checks. In a few-shot prompt the worked examples show the model the exact shape of a good answer, and that demonstration is typically a stronger format signal than a written rule. Here the instruction describes three labelled lines while every example shows a paragraph, and the model follows the examples most of the time. Bringing the examples into line with the instruction removes the contradiction at its source, preserves the tone and length steering the examples were introduced for, and fits the constraints of no extra calls and no tier change.

Why A is correct: Examples are the strongest signal of output format in a few-shot prompt, and these four still demonstrate the retired paragraph shape. Rewriting them removes the conflict, keeps the tone and length control they were added for, and needs no extra calls or model change.

Why B is wrong: Emphasis and placement do make an instruction more salient, so this is a natural first move. It leaves four examples demonstrating the old paragraph format, so the prompt still sends two conflicting format signals and the examples, which show the exact output shape, keep pulling the model back to them.

Why C is wrong: Removing the conflicting examples would probably fix the format, which makes this attractive. The stem says the examples were added because zero-shot updates varied in tone and length, so deleting them trades away a control the team still needs when updating them keeps both.

Why D is wrong: A checking pass is a reasonable compensating control and would catch format failures. The stem rules out extra model calls per update, and a check-and-retry loop treats the symptom while the prompt keeps producing the wrong format most of the time.

See more CCAR-P practice questions, answers explained.

Exam traps in Claude Models, Prompting & Context Engineering

Answers that look right on this material and are not. Each one is a distractor from a different question in the CCAR-P bank for this domain.

Add the 30 examples, because more labelled examples tend to lift accuracy on classification tasks
Why it is wrong: This is tempting because few-shot examples do often improve classification, particularly where categories carry unusual meanings. It is wrong here because the zero-shot prompt already clears the agreed bar on everyday categories, and 30 examples would add input tokens to every request in a service that is already close to its cost ceiling.
Retrieval is returning another patient's record, so the extra medication comes from a mismatched source document.
Why it is wrong: A mismatched record is a real failure mode in clinical retrieval, so it is tempting. It is wrong because retrieval is unchanged and the dose that appears matches one of the prompt's examples, which points at the examples rather than at another patient's record.
Keep the reasoning and send the clinicians' questions through the Message Batches API
Why it is wrong: This is tempting because batch processing lowers cost for large volumes of requests. It is wrong because batches are processed asynchronously for non-urgent work, so a clinician waiting on an answer would see far worse latency, and the unnecessary reasoning tokens would remain.

Examworthy is not affiliated with or endorsed by Anthropic. Original, blueprint-aligned practice material only.