Free CCAR-P Claude Models, Prompting & Context Engineering Practice Questions

The Claude Models, Prompting & Context Engineering domain makes up 13% of the Anthropic CCAR-P exam. Practice free questions covering model selection across the Claude family, prompt design, and context engineering strategies at scale — each with the correct answer and a detailed explanation. The real exam mixes multiple-choice and multiple-response items; these practice questions use a single-best-answer format.

CCAR-P Claude Models, Prompting & Context Engineering Sample Questions with Answers

Sample Question 1 — Claude Models, Prompting & Context Engineering

A software company classifies 180,000 support tickets daily into 25 routing queues. The production SLA is 800 ms, and the minimum acceptable macro F1 score is 0.92. On a representative evaluation set, a faster model achieves 0.935 macro F1 at 420 ms and $0.003 per ticket; a more capable model achieves 0.947 at 1.1 seconds and $0.014 per ticket. Which model strategy should the architect recommend?

  1. A. Use the more capable model because its higher F1 score outweighs the latency and cost differences.
  2. B. Use the faster model because it meets the measured quality threshold, latency SLA, and operating constraints. (Correct answer)
  3. C. Run both models for every ticket and accept a route only when their classifications agree.
  4. D. Alternate between the models evenly so production traffic continues to exercise both capability levels.

Correct answer: B

Explanation: Correct answer (B): The faster model is best because representative evidence shows that it exceeds the required 0.92 macro F1 score while remaining within the 800 ms SLA and costing substantially less. CCAR-P model selection should be driven by demonstrated workload requirements rather than generic capability ranking. In production, right-sizing avoids paying for quality and reasoning capacity that do not improve the defined business outcome. Why the other options are wrong: - Option A: A higher evaluation score can be valuable, but this model violates the explicit latency SLA and provides improvement beyond the stated acceptance threshold at much higher cost. - Option C: Agreement checks can help with uncertain high-risk decisions, but invoking both models for every routine classification would increase cost and latency without a stated requirement for consensus. - Option D: Continued experimentation can detect model differences, but evenly splitting production traffic is not an appropriate steady-state routing strategy when one model already satisfies all requirements.

Sample Question 2 — Claude Models, Prompting & Context Engineering

Twelve business units use Claude to generate weekly operations reports. Each unit copied the original prompt and now maintains its own version. After six months, required risk disclosures are missing in four units, formatting differs across reports, and changes cannot be traced reliably. Units still need distinct terminology and metrics. Which prompt architecture should the technical lead implement?

  1. A. Select the best current unit prompt as the standard and prohibit business-unit-specific prompt changes.
  2. B. Continue separate prompts but use a model to compare their wording and flag substantial differences weekly.
  3. C. Version a shared core prompt and schema, with bounded unit-specific fields and tested override precedence. (Correct answer)
  4. D. Combine every unit's instructions into one comprehensive prompt and let users specify the desired report type.

Correct answer: C

Explanation: Correct answer (C): A versioned core preserves mandatory disclosures and output contracts, while bounded extension points support legitimate unit terminology and metrics. The key constraint is consistency with controlled customization, not complete uniformity. The underlying principle is separating stable shared instructions from scoped variable content, which enables traceable changes, regression testing, and rollback without creating unmanaged prompt forks. Why the other options are wrong: - Option A: A single locked prompt would reduce drift, but it ignores the stated requirement for distinct business-unit terminology and metrics. - Option B: Drift detection would improve visibility, but it leaves duplicated ownership and inconsistent mandatory behavior as the root architectural problem. - Option D: A comprehensive prompt supports all units centrally, but mixing every unit's instructions increases conflicts and context dilution instead of establishing controlled extension boundaries.

Sample Question 3 — Claude Models, Prompting & Context Engineering

A migration assistant converts legacy batch jobs into a deployment pipeline. In testing, some outputs contain sound migration logic but invalid manifest structures; other outputs produce valid manifests that omit required rollback steps. The pipeline must consume artifacts automatically, but failed deployments can interrupt payroll processing. Which design BEST addresses both failure modes?

  1. A. Use one prompt with stronger capitalization and repeated instructions requiring valid manifests and rollback steps.
  2. B. Generate the complete artifact twice and deploy whichever output receives the higher score from another model.
  3. C. Fine-tune output examples for manifest syntax, then allow the deployment pipeline to accept all generated artifacts.
  4. D. Separate analysis and artifact generation, then apply schema and semantic validation with bounded repair or escalation. (Correct answer)

Correct answer: D

Explanation: Correct answer (D): The evidence shows two distinct problems: migration completeness and machine-readable structure. Staging makes the migration plan and rollback requirements inspectable before artifact generation, while deterministic schema and semantic checks prevent unsafe artifacts from entering the payroll pipeline. The governing principle is that reasoning quality and format compliance require different controls; production automation should reject, repair, or escalate failures rather than relying on imperative wording or unvalidated model judgment. Why the other options are wrong: - Option A: Clearer instructions may improve compliance, but repeated imperative wording cannot deterministically guarantee either valid structure or complete rollback logic. - Option B: Generating alternatives can improve selection in some creative tasks, but a model-based score does not provide deterministic validation for payroll-affecting deployment artifacts. - Option C: Additional syntax examples may reduce malformed manifests, but automatically accepting every output leaves missing rollback semantics and residual structural errors uncontrolled.

Sample Question 4 — Claude Models, Prompting & Context Engineering

An enterprise procurement assistant supports negotiations that span several weeks. Its context currently contains every message, retrieved policy excerpt, and tool result. After long sessions, latency triples and the assistant sometimes follows superseded price targets. Current supplier status can change daily, while legal requires citations for policy-based recommendations. Which context strategy is BEST?

  1. A. Compact stale dialogue into structured state, refresh volatile supplier data, and retain authoritative evidence with provenance. (Correct answer)
  2. B. Keep the full conversation but move the newest messages to the beginning so they receive greater attention.
  3. C. Discard all earlier content after each session and ask users to restate the complete negotiation state.
  4. D. Summarize the full conversation into narrative prose and remove all retrieved passages and citation metadata.

Correct answer: A

Explanation: Correct answer (A): Structured compaction removes superseded detail while preserving active goals, unresolved decisions, and other critical state. Refreshing supplier data addresses its daily volatility, and retaining authoritative passages with provenance satisfies the legal citation requirement. The architectural principle is to manage context according to information role and freshness rather than retaining or deleting everything uniformly, which controls latency without sacrificing grounded decision quality. Why the other options are wrong: - Option B: Reordering may emphasize recent messages, but obsolete targets and stale tool results remain in context and continue consuming tokens. - Option C: Starting over controls context growth, but it imposes substantial user burden and risks losing unresolved decisions and legally important evidence. - Option D: Narrative summaries reduce size, but removing source passages and provenance prevents traceable citations and may obscure critical structured facts.

Sample Question 5 — Claude Models, Prompting & Context Engineering

A pharmaceutical research assistant synthesizes internal studies, regulator notices, and licensed journals. The current system places up to 160 potentially relevant documents into each request. Evaluations show that the needed evidence is usually present, but answers sometimes favor repetitive older studies over recent regulator notices. The response SLA is 12 seconds, and every claim must remain traceable. Which change BEST addresses the failure?

  1. A. Preserve all documents but repeat instructions that recent regulator notices must receive the highest priority.
  2. B. Use a larger-context model and include additional studies so the model has more evidence for each synthesis.
  3. C. Select evidence by relevance, authority, freshness, and diversity, preserve provenance, and expand sources progressively. (Correct answer)
  4. D. Summarize every document with a fast model and provide only those summaries without links to original evidence.

Correct answer: C

Explanation: Correct answer (C): The required evidence is already available, so the failure arises from context composition and dilution rather than missing retrieval. Ranking by relevance, authority, freshness, and diversity prevents repetitive older sources from dominating; progressive expansion limits latency, and retained provenance supports traceability. CCAR-P context engineering treats context capacity as a budget to curate, not an instruction to include everything, because excessive competing evidence can reduce grounded synthesis quality. Why the other options are wrong: - Option A: Priority instructions may influence model behavior, but leaving 160 repetitive and conflicting documents in context preserves the dilution causing the failure. - Option B: A larger context window can support broader research, but adding more evidence worsens the identified competition among sources and may violate the latency SLA. - Option D: Summaries could lower token use, but removing links to original evidence violates the claim-traceability requirement and introduces another lossy interpretation layer.

Sample Question 6 — Claude Models, Prompting & Context Engineering

A contract extraction service handles 70,000 requests daily. Every request uses the same 18,000-token policy guide and output schema, but cache hit rates remain below 8%. Engineers construct prompts by interleaving customer metadata, the contract, policy sections, and schema fields in request-dependent order. Output quality is acceptable. Which change should the architect recommend FIRST?

  1. A. Move all customer contracts into a shared cached prompt and append the policy guide to each request.
  2. B. Create a versioned stable prefix for shared guidance and schemas, then append request-specific content consistently. (Correct answer)
  3. C. Upgrade to a more capable model so each request requires fewer extraction instructions and schema examples.
  4. D. Summarize the policy guide separately for every request before submitting the contract extraction prompt.

Correct answer: B

Explanation: Correct answer (B): The shared guide and schema are substantial stable content, but request-dependent interleaving prevents effective prefix reuse. A consistently ordered, versioned stable region directly addresses the low cache-hit rate while isolating volatile customer data and contracts. The principle is to design prompts around stability boundaries so caching can reduce repeated token processing and latency without changing model behavior or output quality. Why the other options are wrong: - Option A: Caching can benefit shared material, but customer contracts are request-specific and may also create privacy and invalidation problems if placed in a common cached region. - Option C: A different model could affect quality or instruction needs, but output quality is already acceptable and model capability does not fix unstable prompt composition. - Option D: Summarization may reduce tokens, but doing it per request adds latency and cost while failing to reuse the unchanged policy guide.

Sample Question 7 — Claude Models, Prompting & Context Engineering

A global operations team is building an incident assistant for 240 services. Each service has runbooks, dashboards, diagnostic commands, and known-failure examples owned by a different team. Most incidents involve one or two services, and guidance changes weekly. Loading the entire catalog makes initial responses take 18 seconds against a 5-second target and occasionally mixes incompatible procedures. Which context design is BEST?

  1. A. Load the complete runbook catalog once per session and instruct Claude to ignore services unrelated to the incident.
  2. B. Fine-tune the weekly runbook catalog into the model so incident requests no longer need operational context.
  3. C. Maintain separate full system prompts for all services and let operators choose the prompt before every incident.
  4. D. Keep a small investigation framework, identify incident scope, and load versioned service modules on demand. (Correct answer)

Correct answer: D

Explanation: Correct answer (D): A small shared framework provides consistent investigation behavior, while on-demand modules limit context to the one or two services normally involved. This directly addresses the 5-second target, conflicting procedures, and weekly guidance changes. The architectural principle is progressive discovery of modular context when the catalog is large and requests are diverse; versioned ownership also supports freshness and rollback without rebuilding a monolithic prompt. Why the other options are wrong: - Option A: A complete catalog ensures availability, but it has already caused unacceptable latency and procedural conflicts that prompt instructions alone do not remove. - Option B: Fine-tuning can encode stable behavioral patterns, but weekly changing operational facts require rapid updates, ownership, and traceability better served by external modules. - Option C: Separate prompts can isolate services, but manual preselection is brittle during cross-service incidents and creates extensive duplicated prompt maintenance.

Sample Question 8 — Claude Models, Prompting & Context Engineering

A software company classifies 600,000 support tickets per day before routing them to specialist queues. Testing on representative production tickets shows that a smaller Claude model achieves 97.2% routing accuracy at 350 ms and $0.004 per ticket. A more capable model achieves 97.8% at 1.1 seconds and $0.019 per ticket. The approved targets are at least 97% accuracy, less than 800 ms, and no more than $3,000 per day. Misrouted tickets are reversible through an existing triage queue. Which model strategy should the architect recommend?

  1. A. Use the more capable model because its higher aggregate accuracy minimizes all downstream operational risk.
  2. B. Use the smaller model because it meets validated quality and latency targets within the daily budget. (Correct answer)
  3. C. Send each ticket to both models and accept the classification when their outputs agree.
  4. D. Use the smaller model initially, then randomly reprocess half of all tickets with the more capable model.

Correct answer: B

Explanation: Correct answer (B): The smaller model is the best choice because representative evaluation shows that it satisfies the explicit accuracy and latency targets while costing about $2,400 per day. The governing principle is to select the least costly model configuration that reliably meets validated workload requirements, rather than maximizing capability without measurable business need. This matters in production because unnecessary model capacity increases recurring cost and may violate service objectives without materially improving reversible outcomes. Why the other options are wrong: - Option A: A more capable model could be appropriate if the smaller model failed important segments, but it exceeds the daily budget and the observed improvement is not required by the approved target. - Option C: Agreement-based ensembles can improve confidence for some workloads, but invoking both models for every ticket increases cost and latency despite the smaller model already meeting the targets. - Option D: Sampling can support quality monitoring, but reprocessing half the traffic adds substantial cost without addressing an identified quality gap or changing routing decisions.

Keep Practicing

Take the 10-question CCAR-P quick-start test across all 7 domains, or return to the CCAR-P practice test hub for 200+ questions, exam details, and a study plan. Drill the other domains: Integration, Solution Design & Architecture, Evaluation, Testing & Optimization, Governance, Safety & Risk Management, Stakeholder Communication & Lifecycle Management, Claude Models, Prompting & Context Engineering, and Developer Productivity & Operational Enablement.