CCAR-P questions and answers — Domain 2: Claude Models, Prompting and Context Engineering

8 practice questions on Claude Models, Prompting and Context Engineering, the domain weighted at 13% of the CCAR-P exam. Every answer explains why the distractors fail — that is the part that transfers to questions you have not seen before.

Question 1

A high-volume system must cut cost without losing quality on its hardest cases. Which arrangement fits?

1. Move everything to the cheapest tier
2. A tiered design in which a light model handles routing, screening and the bulk of tokens, escalating hard cases to a frontier model
3. Move everything to the frontier tier and reduce request volume
4. Alternate tiers by time of day

Answer: B. The executor–advisor arrangement puts most tokens on cheap rates while preserving quality where it matters, and it is defensible because you can point at the escalation criterion. A flat downgrade sacrifices the hard cases; a flat upgrade does not cut cost.

Question 2

Where should a cache_control breakpoint be placed in a prompt containing a large stable reference document and a per-request user question?

1. On the user question, which changes each time
2. On the last block of the stable reference material
3. On every block
4. Caching cannot be used when part of the prompt changes

Answer: B. Caching matches on prefix, so the breakpoint marks how far the identical prefix extends — the end of the stable material, with the volatile question after it. Marking changing content forces a cache write every request; the fact that part of the prompt varies is the normal case, not a blocker.

Question 3

Which is an accurate statement about cache reads at scale?

1. They count toward input tokens per minute like any other input
2. They are exempt from the input-tokens-per-minute limit, so caching raises effective throughput as well as cutting cost
3. They are billed at the same rate as uncached input
4. They invalidate the cache on read

Answer: B. Cache reads are excluded from ITPM and billed at a fraction of base input. That makes prompt caching a capacity lever in a high-throughput system, not only a cost one — a point that appears in scaling questions.

Question 4

An architect proposes solving degraded quality in long agent runs by moving to a model with a larger context window. What is the objection?

1. Larger context windows are not available
2. Attention dilutes across long contexts, so more room lets the problem grow while raising cost per request
3. Larger windows are incompatible with prompt caching
4. The API rejects requests above a token threshold

Answer: B. This is a Foundations trap re-asked at scale, and the answer is unchanged: trim stale tool results, isolate exploration in subagents, offload state, and use context editing. A bigger window addresses the symptom by making room for more of the cause.

Question 5

In clear_tool_uses_20250919, why does clear_at_least exist?

1. To preserve a minimum number of recent tool results
2. To ensure each activation frees enough tokens to justify the cache invalidation the edit causes
3. To delay clearing until a threshold is crossed
4. To protect named tools from clearing

Answer: B. Clearing invalidates the cached prefix at the point of the edit, so a trivial clear pays a cache write for almost no saving. keep preserves recency, trigger sets the threshold, and exclude_tools protects named tools.

Question 6

How should a system prompt be treated in an enterprise deployment?

1. As a configuration string an engineer can edit as needed
2. As a versioned artefact in source control, with an owner and an eval run gating changes
3. As documentation, since it does not affect behaviour deterministically
4. As a secret, excluded from version control

Answer: B. A prompt change is a behaviour change, so it gets the same treatment as a code deployment: versioned, owned, and regression-tested against the eval suite. Treating it as an editable string is how undocumented patches accumulate.

Question 7

Why pin an explicit model version rather than use a floating alias in a production system?

1. Pinned versions are cheaper
2. So behaviour changes only when you decide it does, and the change can be evaluated before it ships
3. Floating aliases are unsupported in production
4. Pinned versions have higher rate limits

Answer: B. A floating alias changes model behaviour without a release, which means an unevaluated change reaching users. Pinning makes the upgrade a deliberate, tested deployment gated by your eval suite. Price and limits are unaffected.

Question 8

What is the examinable justification for a model tier selection?

1. Published benchmark results
2. Evaluation on your own task and data
3. The tier used by comparable organisations
4. The recommendation in the vendor's documentation

Answer: B. Benchmarks describe someone else's distribution. The defensible justification is a measurement against your own cases, which is also what lets you re-check the decision when a new version appears.

← Previous: Questions & answers — Domain 1  |  Next: Questions & answers — Domain 3 →