CCAR-P domain - 13% of the exam

Claude Models, Prompting & Context Engineering

Claude Models, Prompting & Context Engineering is 13% of the CCAR-P exam. These are the objectives it covers, each with practice questions, with every answer explained.

The domain in numbers

40
Practice questions
5
Objectives
9
Easy
31
Medium

Where people go wrong

  • Choosing a tier or thinking setting without matching it to the stage

    The wrong answers set a model tier or thinking setting without matching it to each stage: extended thinking on the editor plug-in, the most capable tier on high-volume document screening, or one middle tier across every turn of a voice call. Stages carry different constraints. Latency-bound turns need the fastest tier, screening already clears its bar there, and the hard judgement step, such as exemptions or order parsing, earns the most capable tier where the budget or a pause can absorb it.

    See it in a sample question
  • Treating step-by-step reasoning as free to cut or reorder

    The wrong answers treat step-by-step reasoning as something to shorten, leave unwritten or leave after the answer at no cost. Halving the working still leaves its delay on a lookup that gained nothing from it, thinking without writing behaves like a direct verdict, and a justification written after the band only rationalises a decision already made. Reasoning improves an answer only when it is written before the answer, and the same added tokens are what lengthen response time.

    See it in a sample question
  • Assuming worked examples teach only format

    The wrong answers treat few-shot examples as a neutral format guide and look elsewhere when output drifts: retrieval, sampling temperature, model caution, or louder emphasis in the instruction. Examples carry their content, their label mix and their layout into the output. Three cardiology summaries leaked a dose into other specialties, five billing labels out of six pulled ambiguous complaints towards billing, and four paragraph examples outweighed a rewritten instruction asking for three labelled lines.

    See it in a sample question
  • Trimming context by position instead of by need

    The wrong answers shrink context by cutting from one end or paraphrasing it: truncating a supplier sheet, writing a free-text summary of exact allergen values, capping retrieval to a few excerpts, or keeping only the most recent encounters. Each drops facts a later step requires. The correct answers choose by need: a structured handoff holding only the fields a stage uses, or retrieval by question type and time window with allergies still carried forward.

    See it in a sample question

Objectives in this domain

What to study

The official documents this domain's practice questions cite most.

Sample questions from this domain

Free sampleClaude Models, Prompting & Context Engineeringmedium

A software company runs two Claude-based review paths. An editor plug-in suggests fixes as developers type and must return its first token within 1.5 seconds. A nightly job reviews every merged pull request that touches concurrency code, and engineers read its findings the next morning. On 200 held-out concurrency pull requests, enabling extended thinking on the nightly reviewer raised the share of seeded defects found from 58 to 81 per cent and added about 40 seconds per review, and the engineering director has approved the higher cost for that job. Where should extended thinking be enabled?

  • AOn both paths, so that developers and the nightly job receive reasoning of the same depth.
  • BOn the editor plug-in only, because interactive suggestions reach developers the soonest.
  • COn the nightly concurrency review only, leaving the editor plug-in to answer without it. Correct
  • DOn neither path, moving the nightly job to the fastest tier to offset the cost of the reviews.
Enable extended thinking where measured reasoning gains matter and latency is slack, and keep it off paths bound by a tight interactive latency budget. Extended thinking lets the model spend additional tokens reasoning before it produces the final answer, which improves results on multi-step problems such as concurrency defects but lengthens time to the answer and raises token spend. The decision is therefore made per path: an offline job with an approved budget and a measured quality gain is where the trade pays, while an interactive path with a 1.5 second first-token budget is where it does not.

Why A is wrong: Consistency across tools is an appealing principle, and the nightly gain suggests the plug-in might improve too. It is wrong because extended thinking adds generation time before the answer, which the 1.5 second first-token budget on the editor path cannot absorb, and the plug-in's quick fix suggestions were never shown to need deeper reasoning.

Why B is wrong: It is tempting to put the best reasoning where users see it first. It is wrong on both counts: the plug-in has the tight latency budget that extended thinking would break, and the measured 23-point gain in defects found belongs to the nightly concurrency review, which this option leaves without it.

Why C is correct: This is correct because extended thinking earns its latency where a task needs multi-step reasoning and the consumer is not waiting: the nightly review has no one reading it until morning, a measured gain from 58 to 81 per cent, and approved cost. The editor plug-in keeps its 1.5 second budget by answering directly.

Why D is wrong: Cost reduction is a reasonable instinct for a job that runs on every merged pull request. It is wrong because the director has already approved the higher cost, and dropping both extended thinking and model capability on hard concurrency reasoning gives up the measured improvement in defect detection that the job exists to deliver.

Free sampleClaude Models, Prompting & Context Engineeringmedium

A hospital trust's ward assistant summarises overnight nursing notes for the morning handover and runs on the most capable model tier. Its p95 response time is 11 seconds against a stated budget of 4 seconds, and logs show nurses abandon 30 per cent of requests before the summary appears. On 400 held-out notes, the current tier scores 95 per cent on the clinical rubric and the fastest tier scores 93 per cent, against an agreed floor of 90 per cent. A clinician proposes keeping the current tier and enabling extended thinking to raise accuracy further. What is the most likely result if that proposal ships as is?

  • AResponse times climb further past the 4 second budget, so abandonment stays high or worsens even if rubric scores rise. Correct
  • BAbandonment falls because nurses trust the more accurate summaries enough to wait for them to finish appearing.
  • CResponse times fall because extended thinking lets the model reach its summary in fewer output tokens overall.
  • DCost per summary falls because more accurate first drafts cut the number of times nurses ask for a regeneration.
When accuracy already clears its floor and latency breaches its budget, adding reasoning effort worsens the binding constraint instead of fixing it. Tier and reasoning choices trade capability against latency and cost. Here both tiers clear the 90 per cent floor, so accuracy is not the binding constraint; the 11 second p95 against a 4 second budget is, and the abandonment data shows its effect. Extended thinking spends extra generated tokens on reasoning before the summary, which pushes latency further over budget and raises cost. The evidence instead supports evaluating the faster tier, which meets the accuracy floor and moves latency in the needed direction.

Why A is correct: Extended thinking adds reasoning tokens before the visible summary, which lengthens time to response on a tier that already misses the budget. Accuracy already clears the floor, so the proposal improves a figure that is not binding while worsening the one that drives abandonment.

Why B is wrong: Trust in output does affect adoption, which makes this appealing to a clinical stakeholder. It is wrong because the logs tie abandonment to waiting for the summary, and both tiers already clear the accuracy floor, so a small accuracy gain does not address the wait.

Why C is wrong: Better reasoning can sometimes tighten a final answer, which makes this sound plausible. It is wrong because extended thinking generates additional reasoning tokens before the answer, so total generation and response time increase rather than fall.

Why D is wrong: Fewer retries can offset a higher per-call cost in some systems, so this is a tempting argument. It is wrong here because the stem reports abandonment, not regeneration, as the problem, and the added thinking tokens raise the cost of every request.

Free sampleClaude Models, Prompting & Context Engineeringeasy

A retail bank routes customer complaints to one of five teams using a prompt with six few-shot examples, five of which are labelled as billing disputes. Since the examples were added, complaints that the operations lead judges ambiguous between fraud and billing go to billing 88 percent of the time, against 41 percent under the previous zero-shot prompt, and the fraud team reports missed cases. The instruction text and model are unchanged. What is the most likely cause?

  • AThe model's safety training makes it reluctant to label a complaint as fraud without explicit evidence of theft.
  • BBilling is the most common real complaint type, so the model is correctly following the bank's own base rates.
  • CThe examples over-represent one label, so the model has learned to favour billing whenever a complaint is unclear. Correct
  • DSix examples exceed what the model can attend to, so it ignores the instruction and copies the first label it sees.
An unbalanced few-shot set biases a classifier towards its majority label on ambiguous inputs, so examples should cover each class and the hard boundary cases. Few-shot examples act as evidence about what answers are expected. When most examples share one label, the model leans towards that label wherever the input itself does not decide the outcome. Rebalancing the set so each team is represented, and including examples that sit on the fraud and billing boundary, removes the skew the examples introduced.

Why A is wrong: Caution about serious labels is a plausible model trait, so this is tempting. It is wrong because the same model routed ambiguous complaints to billing only 41 percent of the time before the examples were added, so the shift follows the prompt change rather than model behaviour.

Why B is wrong: Base rates do matter for ambiguous cases, which makes this tempting. It is wrong because the model has no access to the bank's real distribution; the only skewed signal it received is the example set, and the fraud team's missed cases show the routing is not correct.

Why C is correct: Five of six examples carry the billing label, so the examples teach a prior as well as a format. On clear complaints the content wins, but on ambiguous ones that skew decides the route, which matches the jump from 41 to 88 percent after the examples arrived.

Why D is wrong: Overloading a prompt is a real concern, so this is tempting. It is wrong because six short examples are a small, normal few-shot set, and the observed pattern is a bias towards the majority label on unclear cases, not a uniform copying of one example.

Free sampleClaude Models, Prompting & Context Engineeringmedium

A hospital group's clinical assistant answers clinicians' questions about one patient at a time. Each question currently sends the patient's entire record, which for long-stay patients runs to several hundred pages, and the privacy office has ruled that each request must carry only the information necessary for the question asked. In a review of 400 medication questions, 7 per cent of answers named a drug that had been discontinued years earlier and appeared only in old notes. Allergies recorded in any past encounter must still reach every medication question. Which changes best meet these requirements? Select TWO.

  • ASend only the most recent fifty pages of notes with each request, since older encounters are rarely relevant
  • BRetrieve only the record sections that match the question's type and time window instead of the full record Correct
  • CAdd a system-prompt instruction telling the model to disregard any drug that appears only in older notes
  • DFine-tune a model on the hospital's historical notes so that it learns which medications are still current
  • EAttach the current medication and allergy lists from the source system as labelled, authoritative blocks Correct
Retrieve only what a question needs, and supply safety-critical facts from an authoritative structured source rather than relying on recency or prompt instructions. Sending a whole record breaches minimum-necessary handling and places stale notes in context, where the model can treat an old prescription as current. Retrieving sections by question type and time window removes that noise and the excess data, but a time window alone would lose old allergies; attaching the structured medication and allergy lists from the source system restores exactly those safety-critical facts. Recency truncation, a prompt instruction or fine-tuning each leave either the privacy breach or the allergy gap in place.

Why A is wrong: Tempting because it shrinks the request and removes most of the stale medication mentions. It is wrong because truncation by recency drops allergies recorded in older encounters, which the stem says must reach every medication question, and it still sends pages unrelated to the question asked.

Why B is correct: Correct. Selective retrieval sends what the question needs, which satisfies the privacy office's minimum-necessary ruling, cuts token usage, and removes years-old notes that were steering answers towards discontinued drugs.

Why C is wrong: Tempting because it targets the observed error with a one-line change. It is wrong because the full record is still sent on every request, breaching the minimum-necessary ruling, and an instruction to ignore material that remains in context is a weaker control than not sending it; it could also suppress a valid long-term medication.

Why D is wrong: Tempting because it promises a model that understands the hospital's records. It is wrong because fine-tuning before retrieval work has been tried is over-engineered, it exposes far more patient data to training, and it does nothing to stop the full record being sent on each request, so the privacy ruling is still broken.

Why E is correct: Correct. Structured lists from the system of record give the model current medications directly instead of leaving it to infer them from narrative notes, and they carry every recorded allergy into each medication question regardless of how old the encounter was, closing the gap that time-windowed retrieval would otherwise open.

Other domains in this exam

See also the CCAR-P cert hub, the study guide, and the cheat sheet.

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