Practice exam 3 (final mock) · CCAR-P

Take this one in the last week before your exam, under real conditions: 63 items, 120 minutes, no notes. The domain mix matches Practice exam 2, so you can compare scores directly. Stems are longer and closer to the senior-level judgment calls the real exam favors.

Questions 1 to 21

An insurance broker must compare 200 client policies against each client's stated coverage needs every month and flag gaps for account managers. Results are needed within two days. Which design is best?

A. A real-time agent that account managers chat with for each policy
B. A multi-agent debate over each policy
C. A workflow that extracts each policy to a coverage schema in parallel through batch processing, then compares against needs in code and drafts flags for review
D. A single prompt containing all 200 policies and all client needs A research agent can browse the web, query the internal finance database, and send email to any address. Security asks for the single change that most reduces the risk of data leaving the company. What should the architect recommend?
A. Remove the agent's ability to send email; analysts share results themselves
B. Log every email the agent sends
C. Add a warning about phishing to the system prompt
D. Move the agent to a more capable model

A summarizer on the balanced tier meets quality targets. Volume is about to grow tenfold, and finance is concerned about cost. What is the best first step?

A. Switch all traffic to the fast tier immediately
B. Cut summary length by 80%
C. Stop summarizing low-priority items
D. Evaluate the fast tier on the evaluation set by segment, then route segments that pass to it

An LLM judge agrees with expert reviewers on only 62% of pass/fail decisions. What should the team do next?

A. Use the judge anyway, since it is cheaper than people
B. Refine the rubric with anchored examples, split it into single criteria, require reasoning before the verdict, and re-calibrate
C. Replace experts with the judge
D. Average the judge's score with a random number

An education platform for teenagers adds a Claude study assistant. The main risk is requests for harmful or age-inappropriate content. Which approach is best?

A. Layered controls: a scoped system prompt, input and output safety classification, escalation paths, and monitoring of flagged conversations
B. Rely on the model's default behavior alone
C. Block all questions outside mathematics
D. Ask users to confirm they are adults

The business owner says the system is ready to launch; compliance says it is not. What is the architect's best role?

A. Side with the business owner to keep the date
B. Side with compliance to avoid risk
C. Present results against the agreed launch criteria, identify the specific gaps, and help the accountable owner decide with both parties
D. Launch to a small group without telling compliance

A RAG assistant for maintenance procedures often misses steps that sit near the boundary between two chunks. What is the best fix?

A. Increase the model's maximum output
B. Chunk by procedure or section so each procedure stays whole, with small overlap where sections are long
C. Reduce chunk size to 50 tokens
D. Remove chunk overlap entirely

A retailer wants its assistant to remember each customer's preferences across sessions. Which design is most appropriate?

A. Keep every past conversation in the prompt forever
B. Fine-tune the model on each customer's chats
C. Store preferences in the system prompt for all users
D. A per-customer memory store with consent, user visibility and deletion, strict isolation between customers, and retrieval of only relevant items

Junior engineers are merging Claude-generated code without adding tests. What is the most effective change?

A. Require tests in CLAUDE.md, enforce coverage in CI, and add a review checklist for AIassisted changes
B. Ban Claude Code for junior engineers
C. Ask seniors to rewrite all AI-generated code
D. Increase the model tier

After adding a reranker and enabling extended thinking, p95 latency rose from 3 to 7 seconds and breaches the 5-second SLA. How should the team find what to change?

A. Remove both changes and never use them again
B. Switch to a faster network
C. Use per-step trace timings to see each change's latency contribution, then evaluate quality with each one alone
D. Raise the SLA to 7 seconds

A general customer-support bot has started giving medication dosage advice when asked. What is the best control?

A. Remove the support bot
B. Add a clear scope rule with a safe redirection for medical questions, an output check for medical advice, and test cases for these requests
C. Add a disclaimer to every answer
D. Raise the temperature

A system prompt contains 30 few-shot examples, causing high latency and cost. Evaluation suggests most examples add little. What should the team do?

A. Reduce to a few diverse examples that cover the hard cases, and confirm with evaluation
B. Add more examples
C. Remove all instructions and keep the examples
D. Switch to the most capable tier

A team wants to use a popular open-source MCP server from a public registry with production data. Which TWO steps are required first?

A. Review its code or provenance and pin an approved version
B. Register it behind the gateway with scoped credentials and audit logging
C. Let each developer install it locally with their own admin token
D. Grant it broad scopes so it works in every case

An IT-operations agent will execute runbook actions such as restarting services. Which design is safest while still useful?

A. Give the agent shell access to all servers
B. Let the agent write and run new scripts as needed
C. Expose each approved runbook as a predefined tool; the agent selects and fills parameters, and risky runbooks require human approval
D. Allow actions only after business hours

The implementation team needs an architecture document. What should it contain?

A. A summary of the sales proposal
B. A list of model features
C. Only a diagram
D. Components and interfaces, data flows, non-functional requirements, ADRs, evaluation thresholds, and operational runbooks

After three days, an A/B test shows variant B with a 1.5-point higher resolution rate, but the result is not statistically significant. The planned duration is two weeks. What should the team do?

A. Ship variant B now
B. Continue to the planned sample size and duration before deciding
C. Stop the test and keep variant A
D. Change variant B mid-test

Full prompts containing customer personal data are logged to a general-purpose logging tool that 300 engineers can read. What should the architect do?

A. Redact personal data before logging, restrict access to raw traces, and set a retention period
B. Nothing, because the engineers are employees
C. Turn off all logging
D. Ask engineers not to look at the logs

In which situation is query rewriting before retrieval most valuable?

A. When every query is a single exact part number
B. When the corpus has only five documents
C. When users ask conversational follow-ups such as "what about the second one?" that depend on earlier turns
D. When retrieval is not used

An engineering team will use Claude Code to update configuration across 400 services. What is the best approach?

A. One session that edits all 400 services at once
B. Ask Claude to merge directly to the main branch
C. Skip tests to save time
D. A plan first, then per-service changes in parallel worktrees or batches, with tests per service and human review of each change

After launch, users ask for a feature that would require write access to the finance system. What should happen?

A. Add the tool immediately
B. Take it through intake: discovery, risk review at the right tier, design update with an ADR, and evaluation before release
C. Refuse all new features
D. Let users build it themselves

A classifier must return a label and a short reason in consistent JSON for downstream code. What is the best approach?

A. Define a tool or structured-output schema with the label as an enumeration, use a low temperature, and validate
B. Ask nicely for JSON
C. Parse free text with regular expressions
D. Use extended thinking

Questions 22 to 42

A transit agency wants a trip-planning assistant that accounts for live service disruptions. Which design fits?

A. RAG over a weekly export of the timetable
B. Tools that query the live schedule and disruption APIs, with retrieval only for static content such as fare policies
C. A model fine-tuned on timetables
D. One agent per transit route

Which practice catches quality regressions before users see them?

A. User feedback buttons
B. Monthly business reviews
C. Waiting for complaint volume to rise
D. An offline regression suite that runs in CI on every change

A manager asks for Claude to read performance reviews and recommend which employees to dismiss. What is the best response?

A. Build it as requested with full automation
B. Decline all HR uses of AI
C. Treat it as a high-risk decision about people: the human decides, any AI support is limited and documented, fairness is tested, and legal and HR approve the design
D. Build it but keep it secret from employees

Some agent tasks take up to 20 minutes, but the web front end times out after 60 seconds. What should the architect change?

A. Run tasks as asynchronous jobs with status updates and a notification when done
B. Raise the front-end timeout to 30 minutes
C. Limit every task to 60 seconds of work
D. Use a larger model to finish faster

A partner API returns 50 KB of JSON per call, and tool results are filling the agent's context. What is the best fix?

A. Increase the context window
B. Call the API less often and guess the rest
C. Switch to a faster tier
D. Have the tool return only the fields the task needs, with pagination or summaries for large results

Leadership asks whether the assistant actually caused the drop in handle time, or whether something else did. What is the strongest evidence?

A. A user satisfaction survey
B. A comparison with a control group that did not use the assistant over the same period, using the same measurement method
C. The vendor's case studies
D. The number of conversations handled

Answers drawn from long policy documents often lack reliable citations. Which prompting approach helps most?

A. Ask the model to be accurate
B. Remove document titles to save tokens
C. Wrap each document in tags with its metadata, ask for relevant quotes first, then the answer with citations to those quotes
D. Put the question before the documents

Which TWO observations most clearly show that an evaluation set needs refreshing?

A. Recent production failures are not represented in it
B. It is stored in a spreadsheet
C. The team prefers a different format
D. Traffic has shifted toward new product lines the set barely covers A team wants Claude to review every pull request automatically. What is the best setup?
A. Run Claude Code headless in CI with read-only repository permissions, posting review comments, with humans still approving merges
B. Give Claude permission to approve and merge pull requests
C. Have Claude rewrite every pull request
D. Run reviews only on Fridays

A public-facing assistant is launching. What should be disclosed to users?

A. Nothing; it might reduce use
B. That they are interacting with AI, what it can and cannot help with, and how to reach a person
C. The full system prompt
D. The model's training data

A SaaS company serves many customers from one RAG assistant. What is the strongest way to prevent one customer seeing another's documents?

A. A system prompt rule about confidentiality
B. An output filter for company names
C. Tenant isolation in the data layer: separate indexes, or a tenant filter enforced on every query from the authenticated tenant identity
D. Asking customers not to upload sensitive files

A bank requires that agent processing and data stay inside its own cloud network. Which runtime fits best?

A. Any hosted agent service, regardless of where it runs
B. Developers' laptops
C. A third-party agent marketplace
D. Agents built on an agent SDK and run in the bank's own environment, calling Claude through a deployment that meets its network requirements

Legal asks the architect to explain a design choice. What should the explanation focus on?

A. Data flows, what is stored and where, retention, access controls, and how decisions are reviewed
B. The model's benchmark scores
C. The code structure
D. The cost per token

After a release, the prompt cache hit rate fell from 80% to 10%. What should the team check first?

A. Whether the provider changed prices
B. Whether users asked longer questions
C. Whether the release changed the content or order of the cached prefix, for example by adding a dynamic value near the top
D. Whether the model tier changed its name

An agent often picks the wrong tool. Its tools are named tool1 to tool9 with one-line descriptions. What should the team do first?

A. Add more tools
B. Give tools clear names and descriptions that explain when to use each, with documented parameters and examples
C. Use a larger model
D. Remove the system prompt

A customer asks what personal data the AI system holds about them and how it was used. What must the architecture support?

A. Nothing, because AI systems are exempt
B. A generic privacy notice only
C. Deleting the customer's account
D. Locating that customer's data across conversations, memory, and logs, and producing it under the organization's access-request process

Three product teams each need to integrate with the HR system. What is the best integration approach?

A. One MCP server for the HR system, owned by the HR platform team, registered and governed centrally
B. Each team writes its own direct integration
C. Copy HR data into each team's prompts
D. Give each team a database password

An agent's average cost per task is acceptable, but a few tasks cost 50 times the average. What should the architect do?

A. Ignore the outliers
B. Switch every task to the fast tier
C. Analyze the long-tail traces, then add per-task budgets with early stopping and escalation to a person
D. Remove all tools

Which TWO items belong in an SLA for an internal knowledge assistant?

A. 100% answer accuracy
B. A p95 latency target
C. A quality floor measured by sampled grading
D. Unlimited throughput

Answers mention product features that do not exist, although retrieval returns the correct documents. Which metric should be added to the evaluation suite?

A. Recall at k
B. Groundedness (faithfulness of claims to the provided sources)
C. Tokens per second
D. Cache hit rate

The platform team publishes a "golden path" template for internal assistants. What is the main benefit?

A. It removes the need for evaluation
B. It lets teams skip security review entirely
C. It locks every team into one prompt
D. Teams ship faster with retrieval, guardrails, evaluation, and observability already in place, and approval is quicker because the controls are known

Questions 43 to 63

An operations agent needs to use the cloud provider's command-line tool. How should this be set up?

A. Give it the CLI with administrator credentials so nothing fails
B. Run it in a sandbox with an allowlist of permitted commands, scoped credentials, and approval for commands that change resources
C. Let it install any CLI it wants
D. Use the CLI only through the model's memory of its syntax

A business unit wants to "replace the contact center with AI agents" in a single launch. What should the architect recommend?

A. A single big-bang launch to all customers
B. Wait until the technology is perfect
C. Replace staff first, then build the system
D. A phased plan: start with high-volume simple intents, keep a human handoff, measure, and expand on evidence

Injection resistance is a release metric. A new release raises the red-team injection success rate from 1% to 6%. What should happen?

A. Block the release until the cause is found and the rate is back within the threshold
B. Release it, because 6% is still low
C. Release it and monitor
D. Remove the metric

After migrating to a newer model, outputs are much longer than before, though accuracy improved. What should the team do?

A. Roll back permanently
B. Accept the longer outputs
C. Adjust the output instructions (length, format) for the new model and re-run the evaluation suite
D. Truncate outputs in code at a fixed character count

A requirement says every response must include an order number in the format ORD-#####. Which grading method fits best?

A. An LLM judge
B. A code-based check such as a regular expression
C. Weekly human review
D. User feedback

Retrieval latency is high because the system queries six indexes one after another. What is the best fix?

A. Remove four indexes
B. Use a larger model
C. Cache all answers forever
D. Query the indexes in parallel, or route to the relevant ones using metadata, then merge and rerank

Discovery interviews show that regions define a "resolved ticket" differently. What should happen before success metrics are built?

A. Agree one definition with the accountable owners and record it
B. Use each region's definition and average them
C. Let the model decide what counts as resolved
D. Skip the metric

For simple create-read-update-delete code changes, a team proposes three agents (planner, coder, tester) that pass their full context to each other. What is the best assessment?

A. Ideal; more agents means better quality
B. Fine, as long as each agent uses the most capable tier
C. Over-engineered: a single agent with tests, or a simple workflow, is cheaper and more reliable for this task
D. It should use five agents instead

A team wants to use Claude on confidential merger documents. What is the most appropriate configuration?

A. Any consumer chat app employees already use
B. An approved enterprise deployment with appropriate retention terms (zero data retention where eligible and required), access restricted to the deal team, and audit logging
C. Email the documents to a personal account and paste them in
D. No controls, because the documents are internal

The team wants to switch to a newer embedding model. What does the change require?

A. Re-embed the corpus with the new model and compare retrieval metrics on the labeled query set before switching traffic
B. Nothing; embeddings from different models are interchangeable
C. Re-embed only new documents
D. Switch the query embeddings only

A Claude-powered service has no dashboard yet. What should a minimum dashboard include?

A. CPU and memory only
B. Number of deployments
C. Total monthly invoice only
D. Volume, latency percentiles, errors and rate limits, cost by feature, cache hit rate, sampled quality scores, and escalation rate

A long-running agent forgets constraints the user gave at the start of the session. What is the best fix?

A. Ask the user to repeat the constraints every hour
B. Keep the constraints in a persistent place the agent always sees (system prompt or progress file), and make compaction preserve decisions and constraints
C. Use a smaller context window
D. Remove the constraints

Which TWO actions best manage expectations at project kickoff?

A. Share evaluation-based targets and known limitations
B. Promise that the system will learn and improve on its own
C. Agree the human-review policy and the escalation path
D. Avoid discussing possible errors

Which metric best shows whether Claude Code improves onboarding for new engineers?

A. Number of prompts per day
B. Lines of code generated
C. Time to first merged, production-quality change
D. Number of installed extensions

An executive asks for the cost of running a proposed assistant. What should the estimate include?

A. Model tokens, infrastructure, observability, human review, and support, set against measured value versus the baseline
B. Model token cost only
C. The license fee only
D. Nothing until after launch

An assistant must work with twelve SaaS tools, and teams want all twelve tool schemas loaded on every request. What is the better design?

A. Load all twelve always
B. Role-based allowlists for what each agent may use, plus progressive discovery so only relevant tool definitions are loaded per task
C. Remove all but one tool
D. Ask users to choose tools manually each time

A load test before a seasonal peak shows rate-limit errors at twice expected volume. What should the team do?

A. Hope traffic stays low
B. Remove the rate-limit handling
C. Request higher limits ahead of time, add queueing with backoff, confirm fallbacks, and re-test
D. Switch off the assistant during the peak

What should a post-incident review for an AI system produce?

A. A list of people to blame
B. A new system prompt only
C. Nothing, if the incident is resolved
D. A blameless account of causes, new test cases, design fixes where possible, and owners with dates

Executives and engineers need different summaries of the same incident data. What is the best prompt design?

A. Shared base instructions plus separate, versioned audience modules for each summary type
B. One prompt that tries to satisfy both audiences at once
C. Let each reader write their own prompt
D. Send engineers' summaries to executives

A Claude assistant is being retired. Which set of steps is most complete?

A. Turn it off
B. Notify users and offer alternatives, delete or archive data per retention rules, revoke credentials and integrations, and archive documentation
C. Delete the documentation first
D. Leave it running in case someone needs it

The observability vendor stores traces outside the region where customer data must remain. What should the architect do?

A. Ignore it, because traces are not customer data
B. Stop all monitoring
C. Keep traces in-region (or use an in-region option), redact personal data before export, and document the data flow
D. Move customers to another region

Practice exam 3: answer key and rationales

Domains follow the same item numbering as Practice exam 2, so the domain scoring table above applies here too.

# Answer Domain Why

1 C 1 Fixed, parallel, non-urgent work: workflow plus batch, comparison in code. 2 A 3 Removing the outbound channel breaks the exfiltration path.

3 D 2 Evaluate per segment before routing to a cheaper tier.

4 B 4 Low judge agreement means the rubric needs work and recalibration.

5 A 5 Layered safety controls for a vulnerable audience.

6 C 6 Evidence against agreed criteria; the accountable owner decides.

7 B 3 Structure-aware chunks keep procedures whole.

8 D 1 Memory is designed explicitly, with consent and isolation.

9 A 7 Guidance plus enforcement in CI plus review habits.

10 C 4 Trace each step's latency, then decide with evaluation.

11 B 5 Scope rule, output check, and test cases for out-of-scope advice.

12 A 2 A few diverse examples; confirm with evaluation.

13 A, B 3 Review and pin, then govern behind the gateway.

14 C 1 Predefined runbook tools with approval for risky actions.

15 D 6 Implementation guidance needs interfaces, NFRs, ADRs, thresholds, runbooks. 16 B 4 Do not stop early on a non-significant result.

17 A 5 Redact, restrict, and set retention.

18 C 3 Rewriting resolves conversational references before search.

19 D 1 Decompose, parallelize safely, test, review.

20 B 6 New high-impact capabilities go back through intake and risk review.

21 A 2 Schema with an enumeration, low temperature, validation.

22 B 1 Live data needs live tools, not a stale index.

23 D 4 Offline regression gates catch problems before release.

24 C 5 Decisions about people stay human, with fairness and legal review.

25 A 1 Long work belongs in asynchronous jobs.

26 D 3 Tools should return only what the task needs.

27 B 6 A control group isolates the assistant's effect.

28 C 2 Tagged documents and quote-first answering improve citations.

29 A, D 4 Unrepresented failures and traffic shifts mean the set is stale.

30 A 7 Headless review with read-only rights; humans merge.

31 B 5 Transparency: AI disclosure, scope, human route.

32 C 3 Isolation belongs in the data layer.

33 D 1 Network and data requirements favor running agents in your own environment. 34 A 6 Legal needs data flows, retention, access, and review.

35 C 4 Cache misses usually mean the prefix changed.

36 B 2 Tool names and descriptions are prompts; make them clear.

37 D 5 Access requests must reach every store of personal data.

38 A 3 Shared integrations belong in a governed MCP server.

39 C 1 Study the long tail, then budget and escalate.

40 B, C 6 Latency targets and quality floors are measurable SLA terms.

41 B 4 Groundedness measures fabrication against sources.

42 D 7 Golden paths make the safe way the fast way.

43 B 3 CLIs run sandboxed, allowlisted, scoped, with approvals.

44 D 1 Phase by intent and evidence, with human handoff.

45 A 5 A breached safety threshold blocks release.

46 C 2 Re-tune prompts for the new model and re-evaluate.

47 B 4 Deterministic requirements get deterministic checks.

48 D 3 Parallelize or route retrieval, then merge.

49 A 6 One agreed definition before measurement.

50 C 1 Simple tasks do not need multi-agent systems.

51 B 5 Sensitive data needs an approved deployment and access controls.

52 A 3 Changing the embedding model means re-embedding and re-measuring.

53 D 4 LLM dashboards need quality and cost signals, not only infrastructure. 54 B 2 Persist constraints; make compaction keep them.

55 A, C 6 Honest targets and clear review and escalation set expectations.

56 C 7 Measure outcomes for new engineers, not activity.

57 A 1 Full cost against measured value.

58 B 3 Allowlists for safety, progressive discovery for efficiency.

59 C 4 Plan capacity, queue, back off, fall back, re-test.

60 D 5 Blameless review with tests and design fixes.

61 A 2 Modular prompts per audience on a shared base.

62 B 6 Retirement covers users, data, credentials, and records.

63 C 3 Residency applies to traces too.

If you score 51 or more on both Practice exam 2 and Practice exam 3, with every domain at or above its "ready at" level, you are in a strong position to book the exam.

Appendices

A. Architecture review checklist

Use this in real design reviews and as a last-week revision sheet.

Outcome, baseline, and target metric written down
Simplest pattern chosen, with the reason recorded in an ADR
Model tier per step justified by evaluation results

Static prompt content first, caching enabled where traffic repeats Every tool needed by the role; write tools narrowed or gated

Authorization enforced per user in tools and retrieval

Untrusted content isolated; outbound paths allowlisted or approved Retrieval strategy matches data shape; index freshness monitored Evaluation set from real traffic, with code, model, and human grading Regression suite gates prompt, model, and config changes

Traces with correlation IDs, versions, tokens, latency, cost

Human oversight sized to risk tier

Compliance obligations mapped (GDPR, HIPAA, FedRAMP, sector rules)

SLA with quality floor and fallback agreed with the owner

Handoff pack: runbooks, dashboards, owners, migration plan

B. Glossary

Term Meaning

A/B test Comparing two variants on live traffic with randomized assignment ADR Architecture decision record: context, options, decision, consequences Agent A system where the model chooses its own steps and tool calls in a loop Agent-to-agent Integration where one agent delegates work to another agent with its own tools and policies Augmented LLM A single model call enriched with retrieval, tools, or memory BAA Business associate agreement, required for PHI under HIPAA

Batch processing Asynchronous, discounted processing of many requests, with results returned later Canary release Sending a small share of traffic to a new version before full rollout Term Meaning

Capability bloat Tools or permissions beyond what a role needs

Chunking Splitting documents into retrievable pieces

Circuit breaker A mechanism that stops calling a failing dependency for a period CLAUDE.md A file of project knowledge and conventions that Claude Code loads into sessions Compaction Summarizing a long context to free space while keeping key information Confused deputy A privileged component tricked into acting for a less privileged user Contextual retrieval Adding a short context line to each chunk before indexing Correlation ID An identifier linking all calls in one request or task for tracing DPIA Data protection impact assessment under GDPR

Embedding A numeric vector representing text meaning, used for semantic search Evaluator-optimizer A pattern where one call drafts and another critiques until criteria pass Extended thinking A mode where the model reasons at length before answering, within a budget FedRAMP US federal program authorizing cloud services for government use Few-shot prompting Including a few worked examples in the prompt Golden path A supported, documented way to build a common type of system Groundedness The degree to which claims are supported by provided sources Guardrail A control that constrains inputs, actions, or outputs

Hook A script Claude Code runs automatically at a lifecycle event Human in the loop A person reviewing or approving model output or actions Hybrid search Combining keyword and vector retrieval

Idempotency A repeated call has the same effect as one call

Lethal trifecta Private data, untrusted content, and an external channel in one agent LLM-as-judge Model-based grading against a rubric

Managed settings Organization-wide Claude Code settings users cannot override MCP Model Context Protocol, an open standard for connecting models to tools and data MCP gateway A central layer for auth, policy, rate limits, and audit across MCP servers Orchestrator-workers A lead model splits work dynamically and delegates to workers p95 latency The response time that 95% of requests are faster than PHI Protected health information under HIPAA

Progressive discovery Loading tools, skills, or documents only when needed Prompt caching Reusing an unchanged prompt prefix to cut cost and latency Prompt chaining Splitting a task into sequential prompts with checks between Prompt injection Instructions hidden in input or retrieved content that hijack the model RACI Responsible, Accountable, Consulted, Informed: a decision-rights matrix RAG Retrieval-augmented generation

Recall at k Share of queries whose relevant item appears in the top k results

Appendices

Term Meaning

Red teaming Adversarial testing to find failures and vulnerabilities Reranking A second pass that reorders retrieved chunks by relevance Routing Classifying a request and sending it to a specialized path or model Skill A packaged folder of instructions and resources loaded on demand SLA Service level agreement: committed availability, latency, and quality Structured output Model output constrained to a schema

Subagent An agent launched by another agent for a subtask, with its own context Token passthrough Forwarding a client's token to downstream services; an anti-pattern Tool use The model calling functions defined by the application

ZDR Zero data retention arrangement with the provider

C. 30-day study plan

About one hour a day on weekdays and two on weekends, roughly 40 hours in total.

Days Focus Output

1 to 2 Exam guide, Chapter 1, take Practice exam 1 cold Baseline score by domain 3 to 7 Chapter 4: Integration Diagram one RAG pipeline and one MCP topology

8 to

Chapter 2: Solution Design Write two ADRs for a real or invented system 12 to

Chapter 5: Evaluation Build a 30-case eval set with a rubric

16 to

Chapter 6: Governance Threat-model one agent with the risk catalogue

19 to

Chapter 7: Stakeholders and Lifecycle One-page executive brief and a handoff pack outline

22 to

Chapter 3: Models and Context Measure caching and model-tier trade-offs on a small task

25 Chapter 8: Developer Productivity Write a team Claude Code settings and CLAUDE.md template

26 to

Chapter 11 capstones Answer all cases before reading the model answers; do the labs in Chapter 12

28 Practice exam 2, full length, timed at 120 minutes Target 51 of 63 or more 29 Review weakest domain and the checklist Notes on every missed item 30 Rapid review (Chapter 13), logistics check, rest; Practice

exam 3 a few days before the exam if not done

Ready

D. One-page domain summaries

Read these the night before the exam.

Domain 1: Solution Design & Architecture (17%)

Start from a measurable outcome and baseline.

Pattern ladder: single call, augmented LLM, workflow, agent, multi-agent. Climb only as far as the task needs.

Every design has input controls, processing, validated output, and a feedback loop. Multi-agent only for parallel, wide, or isolated-context work; subagents return summaries. Hard limits on agents: turns, tools, spend, time. Idempotent side effects.

The SLA can rule a pattern out; long work goes asynchronous.

Platform choice follows constraints: commitments, residency, networking, feature availability. Domain 2: Models, Prompting & Context Engineering (13%)

Choose tiers per step with evaluation evidence; route and escalate.

Prompts: role, delimited context, rules with reasons, output contract, behavior when unsure. Few-shot: diverse examples; chain of thought or extended thinking only where it pays. Static first, dynamic last; cache the stable prefix.

Context is a budget: prioritize, compact, isolate, load on demand (Skills, tool search). Pin model versions; migrate with evaluations and canaries.

Domain 3: Integration (19%)

Remove, then narrow, then gate, then monitor.

The agent's rights never exceed the user's; authorization lives in tools and data. No token passthrough; short-lived scoped credentials; secrets never in context. MCP for shared reusable integrations with a registry and gateway; direct tools for one-offs; CLIs in sandboxes; agent-to-agent for partner-owned agents.

Retrieval fits data: hybrid for identifiers, SQL tools for structured aggregates, context plus cache for small corpora.

Chunk by structure, add context to chunks, filter by permissions and freshness, rerank. Trace everything with correlation IDs and versions; redact before logging.

Domain 4: Evaluation, Testing & Optimization (16%)

Define metrics per dimension: accuracy, groundedness, latency, cost per success, safety, security. Eval sets from real traffic, stratified, expert-labeled, with a frozen holdout. Code checks, calibrated LLM judges, sampled human review.

Regression suite gates every change; one variable at a time; canary and A/B.

Appendices

Diagnose from the symptom: data change points to retrieval; hard cases point to model mismatch. Optimize in order: caching, batch, routing, trimming context, limiting output, streaming, parallelism.

Domain 5: Governance, Safety & Risk (14%)

No prompt makes a model immune to injection; limit the blast radius (lethal-trifecta test). Layered guardrails: input, model, tools, output, process.

Oversight sized to risk; humans decide critical and regulated outcomes; make review meaningful. GDPR: lawful basis, minimization, rights, DPIA, transfers, automated decisions. HIPAA: BAA and eligible configuration. FedRAMP: authorized environment.

Fairness testing by segment; transparency to users; audit trails.

Domain 6: Stakeholder Communication & Lifecycle (14%)

Discovery before design; problem statements with user, baseline, target, constraints, owner. ADRs for decisions; message tailored to each audience.

Honest numbers about non-determinism; SLAs with quality floors and fallbacks. When results fall short: data, cause, options, owner decides.

Handoff with runbooks, eval suite, dashboards, owners, and a shadow period.

Lifecycle: discovery, design, build, launch, handoff, monitor, iterate, retire. Domain 7: Developer Productivity & Operational Enablement (7%)

Managed settings for rules that must never be broken; project settings and CLAUDE.md for team conventions.

Hooks and permissions enforce; instructions guide.

Approved MCP servers with scoped credentials; read-only production by default. Headless use in CI with tight permissions and human merge.

Measure delivery outcomes, not activity.

E. Exam FAQ

Do I need to pass Foundations first? No. The official exam guide lists no mandatory prerequisite, but it recommends substantial architecture and production Claude experience. How many questions and how long? 63 multiple-choice and multiple-response items in 120 minutes, according to the exam guide.

What score passes? A scaled score of 720 on a 100 to 1,000 scale. The score report also shows percent correct by domain, but those percentages do not decide the result.

Are there scenario sets like on Foundations? Third-party guides describe Professional items as standalone rather than grouped into scenarios. Expect each item to carry its own short scenario. Is there a penalty for guessing? The exam guide does not mention one. Answer every item. Can I use documentation during the exam? Reported as closed book; check the candidate rules when you book.

How long is the certification valid? 12 months from the award date. Renewal arrangements are described by Anthropic's partner FAQ and third parties; confirm the current process in the Partner Academy.

What if I fail? Use the domain breakdown to target study, observe the retake waiting periods, and re-sit. Retake fees and limits are set in the candidate policies.

How current must my product knowledge be? The exam tests architecture judgment more than feature trivia, but you should know current capabilities such as prompt caching, batch processing, tool use, MCP, Skills, extended thinking, and Claude Code configuration.

F. Sources

Exam facts were checked on 29 October 2026. Anthropic's official exam guide and partner FAQ are the authority; everything else is secondary.

Claude Certified Architect – Professional Exam Guide v1.0 (official PDF): items, timing, scoring, fee, domains, task statements, sample questions Credly badge: Claude Certified Architect – Professional: issuer, skills, earning criteria Tutorials Dojo CCAR-P study guide: domain topics; links to the official guide and Partner Academy page

Training Camp: Claude certifications explained: eligibility, delivery, discounts, renewal (third party)

Udemy blog: Claude certifications guide: prices across the four exams (third party) Claude Certification Guide: CCAR-P blueprint: domain weights (independent, not affiliated with Anthropic)

Retake rules and Japanese-language availability come from Anthropic's partner certification FAQ as reported by the companion CCAR-F book; confirm them on the Partner Academy before you book.

Appendices