No account needed
Free CCAR-P mock exam, timed, with a reason for every option
63 questions written for the Claude Certified Architect - Professional blueprint, sat at the official pace of about 1.9 minutes an item. No sign-up, nothing saved, nothing sent anywhere.
Want more? The post CCAR-P architecture scenarios explained works 10 more items option by option, and the CCAR-P study guide shows 6 more with their keys: 79 distinct items with this paper, none behind an account.
Three ways to sit it
Start the CCAR-P mock
Answer in this tab and get a raw score by domain with the key and the reason each option fails. Every mode scores the same way and shows the same review afterwards. The clock is the only difference, and it is the difference most candidates underestimate. Or work through the paper below one question at a time, checking each as you go.
Full, timed
All 63 items in 120 minutes, about 1.9 minutes an item, the official pace. The sitting submits itself when the clock reaches zero.
Quick, timed
30 items drawn evenly across the 7 domains, in 57 minutes at the same pace. A different draw each time.
Untimed review
Every item with no clock. Flag what you are unsure of, submit when you are done, and read the reasons.
Filter the paper by domain
- All domains 63
- D1 Solution Design & Architecture 11
- D2 Claude Models, Prompting & Context Engineering 8
- D3 Integration 12
- D4 Evaluation, Testing & Optimization 10
- D5 Governance, Safety & Risk Management 9
- D6 Stakeholder Communication & Lifecycle Management 9
- D7 Developer Productivity & Operational Enablement 4
Answer as you read
The CCAR-P paper: every question
63 items, each tagged with its official domain. Work through the paper one question at a time: pick an option to check it, or ask for the key, and read why each option holds or fails. Nothing is saved. For a scored sitting against the clock, use the start modes above.
Checked 0 of 63, 0 correct
Sit it timed insteadA parcel carrier receives 9,000 customer emails a week, mostly asking where a parcel is. Tracking status changes every few minutes. Which Claude-based design best meets a fifteen-minute reply target?
Select the best answer.
A scheduling agent carries 31 tools, some that cancel referrals or export patient lists. The role needs six of them, none that cancel or export. Wrong-tool selection is 14 percent. What should you do first?
Select the best answer.
An article classifier must reach 95 percent accuracy on a labeled set. The top tier scores 97.1 percent; the fastest tier scores 96.4 percent at one twelfth the cost. What should you do?
Select the best answer.
Every auto claim goes through the same five steps: extract, classify damage, check coverage, estimate a reserve, and route. Regulators must be able to reproduce each routing decision. Which architectural pattern fits best?
Select the best answer.
An internal MCP server gives 600 branch staff account lookups. It reads any customer through one service account, choosing the customer from the account number typed in chat. What closes the security gap?
Select the best answer.
An orchestrator sends three subagents to review supplier quotes. Each returns its full transcript, so the orchestrator holds 180,000 tokens and its recommendation contradicts the subagents' findings. What should change?
Select the best answer.
A billing assistant sometimes agrees to waive late fees when customers insist, though only a supervisor can approve a waiver and the assistant has no waiver tool. How should the system prompt change?
Select the best answer.
Gate agents ignore answers slower than two seconds; wrong answers are usually caught at the booking screen. At p95 the fast tier takes 1.6 seconds, the more accurate large tier 3.2. Which tier fits?
Select the best answer.
Clinicians edit and sign discharge summaries that Claude drafts. About 41 percent need heavy rewriting, but nobody knows what changes, and no step may add clinician time. What closes the loop?
Select the best answer.
Contracts are indexed as fixed 512-token chunks with no metadata. For questions about one client's agreement, 31 percent of top chunks come from a different contract, and many split a clause. What should change?
Select the best answer.
A template pastes each supplier's document into the system prompt. One document said to ignore the format and reply in French, and 340 records failed to load. Which two changes should you make? (Select two.)
Select exactly 2. 0/2 selected.
One prompt reads a 60-page grant application, checks 14 eligibility rules, scores six criteria and drafts a letter. The eligibility check is right only 58 percent of the time. How should the work be decomposed?
Select the best answer.
A shopping assistant logs every full prompt and response from two million daily sessions. Logging now costs more than the model, and a broken promotion rule went unnoticed for two days. What should change?
Select the best answer.
An associate asks whether any of a target company's 3,000 agreements contain change-of-control clauses affecting its loans. Which documents matter depends on earlier findings, and a partner reviews every conclusion. Which pattern fits best?
Select the best answer.
A troubleshooting agent runs up to 40 turns, and each diagnostic tool returns about 8,000 tokens of router logs that stay in the conversation. After turn 20 it forgets early constraints. What should change?
Select the best answer.
A claims agent has 27 tools, including four overlapping search tools and run_query, which runs any SQL. Wrong-tool selection is 11 percent, and security has flagged run_query. What should you do?
Select the best answer.
A monthly safety report must synthesize 2,400 incident reports. Putting them all in one prompt overflowed the context window, and a subset produced trends nobody could trace to a report. Which approach fits best?
Select the best answer.
Engineers search 90,000 documents indexed with dense embeddings. Most queries contain an exact part number or fault code, and for those the top result often covers a similar but different code. What should change?
Select the best answer.
A planning assistant sends an 11,000-token system prompt on every request, with caching enabled. The prompt starts with the current time and the applicant's reference, and the hit rate is 3 percent. What fixes it?
Select the best answer.
Five chained agents write product descriptions: tone, keywords, banned phrases, translation and review, then editors approve each. Costs are nine times one call, and p95 is 40 seconds against a six-second budget. What should change?
Select the best answer.
A public assistant lets anyone report a fault. Its tools file work orders through a privileged service account and read the staff roster to name the responsible crew. What is the best change?
Select the best answer.
On 900 rebooking cases, the fast, middle and top tiers score 78, 91 and 93 percent. The floor is 90 percent, and only the top tier exceeds the eight-second p95 budget. Which tier fits?
Select the best answer.
An operations chief sponsored a mortgage-document system to lower cost per file while keeping a five-day turnaround. Six weeks in, cost per file is down 40 percent. How should the architect present results?
Select the best answer.
Seven teams build AI tools that need the same student-record lookups. Each wrote its own API wrapper, three handle the consent flag differently, and the registrar cannot see who accesses records. Which integration fits?
Select the best answer.
Your team sorts guest reviews into twelve house-specific labels. A zero-shot prompt with definitions scores 74 percent, and most errors fall between neighboring labels. You have 300 labeled reviews. What should you try first?
Select the best answer.
Claude reads adverse-event emails and writes records straight into a safety database that regulators inspect. In a pilot, 7 percent of records had a wrong or missing field. Which two design changes fit best? (Select two.)
Select exactly 2. 0/2 selected.
A medical-information assistant's unhelpful-answer rate rose from 4 to 13 percent in six weeks. Errors, latency and tokens are flat, the model is unchanged, and the corpus is refreshed weekly. What monitoring is missing?
Select the best answer.
You are designing two integrations: a nightly job that classifies 200,000 delivery notes with no user waiting, and a partner's scheduling agent negotiating slots with yours across company boundaries. Which two choices fit? (Select two.)
Select exactly 2. 0/2 selected.
Fourteen assistants each copy the same 3,000-token block of style rules and disclaimers. One disclaimer change took three weeks, and two teams shipped the old text. What should you do?
Select the best answer.
A lender's board wants generative AI. Each application takes 2.5 hours of re-keying figures from scanned documents and 20 minutes writing a memo. A regulated model makes the credit decision. Where should Claude help?
Select the best answer.
An operations agent loads all 58 tool definitions, about 42,000 tokens, on every request, though a task uses three to five. Tool-selection errors are 12 percent. Which context strategy fits best?
Select the best answer.
A contract promises brokers a response within 4 seconds. The dashboard reports a 2.3-second mean and marks the SLA met, yet one broker logged 40 slow-response complaints. What should be measured?
Select the best answer.
An admissions assistant passed 40 tests written by the developer of its prompt, yet 18 percent of sampled live answers were wrong. The registrar can spare four hours. What should you build?
Select the best answer.
Two days into a two-week test, a prompt variant resolves 71 percent of support chats and control 68 percent, about 300 each. Traffic peaks on Mondays; the test began Wednesday. What should you do?
Select the best answer.
Clinicians sign off every summary draft, and some drafts now name medications absent from the notes. That week a smaller tier and wider retrieval (6 to 20 chunks) shipped together, untested. What should come first?
Select the best answer.
A tariff-code classifier scores 94 percent on its evaluation set, but a broker says a third of codes for Portuguese and Spanish descriptions are wrong. Every evaluation item is in English. What does this show?
Select the best answer.
A weekly job regenerates 70,000 product descriptions with synchronous calls, each carrying the same 6,000-token style guide. Nobody reads the output until Monday. Finance wants cost cut by a third. What fits best?
Select the best answer.
A chat assistant's p95 is 9 seconds against a 4-second target. Tracing shows generation steady near 2 seconds, while an identity lookup tool sometimes takes 7. What should you do?
Select the best answer.
A manual-search assistant logs only HTTP status and total latency. When answers grew vaguer over several weeks, nobody could tell whether the prompt, model or indexed manuals had changed. What logging fits best?
Select the best answer.
A law firm's assistant extracts clauses from client contracts and reports only 96 percent extraction accuracy. A client asks how it is tested for safety and security before each release. Which two metrics fit? (Select two.)
Select exactly 2. 0/2 selected.
You must compare 500-token and 1,200-token chunking for a fare-rules assistant. You have 600 labeled questions, and no customer may knowingly get a worse configuration during peak season. How should you compare them?
Select the best answer.
A member assistant can waive fees up to $50 through a tool. Pasted text told it to waive a fee on an ineligible account, and it did. Which control fits best?
Select the best answer.
Policy forbids sending a resident's national ID number to the model provider. Residents type it in about 15 percent of chats, always in the same nine-character format. Which control fits best?
Select the best answer.
An agent downloads PDFs from unknown suppliers, extracts pricing, drafts a comparison table and can email it to committee members. The risk register must be signed off first. Which risk belongs at the top?
Select the best answer.
A claims assistant recommends approve, request evidence or deny on 3,000 claims weekly. A wrong denial leaves a customer without a car for weeks; a wrong small approval costs little. Where should review sit?
Select the best answer.
A pharmacy team approves drafted medication-change letters in a median of 9 seconds; 98 percent pass unchanged. The screen shows only the draft and two buttons. Two wrong-dose letters got through. What should change?
Select the best answer.
A prescription assistant runs under a business associate agreement. Before a HIPAA audit, you find its observability platform, whose vendor has signed no such agreement, stores full prompts with patient data. Which conclusion is right?
Select the best answer.
A student-support assistant is extending to EU applicants. Each request sends the full applicant record, including disability and nationality, and every transcript is kept for evaluation. Which two controls address GDPR most directly? (Select two.)
Select exactly 2. 0/2 selected.
A loan pre-screening assistant drafts recommendations that officers confirm. Protected attributes are absent, but ZIP code and employer are present. A risk committee asks what shows the assistant treats groups fairly. What should you provide?
Select the best answer.
After three months, maintenance engineers accept 97 percent of drafted fault diagnoses without opening the logs, and two incidents trace to wrong diagnoses accepted unchanged. Which failure mode is this, and what mitigates it?
Select the best answer.
An operations chief wants an assistant to answer all 2,500 daily customer emails and asks for a proposal in two weeks. The emails span bookings, disputes, invoices and damage claims. What should discovery establish first?
Select the best answer.
A finance chief must choose between calling the Claude API directly and using it through a cloud platform the company already uses. The decision memo lists only each route's benefits. What should it add?
Select the best answer.
An editor proposes that 99 percent of suggested headlines be "acceptable" and that suggestions return within 2 seconds. The pilot measured a lower acceptance rate, and nobody has defined "acceptable". What should the architect do?
Select the best answer.
Dashboards show latency, errors, cost and a weekly evaluation score. Over six weeks the score fell from 88 to 79 with errors flat, and nobody acted until a plant manager complained. What governance is missing?
Select the best answer.
An inherited assistant came with a diagram, an API reference and a runbook. A new engineer raised the chunk size and unknowingly broke a control that kept account data out of retrieval. What was missing?
Select the best answer.
Two months after launch, the partner has left, the model faces deprecation, the catalog has changed and the evaluation score fell from 90 to 84. Nobody owns the assistant. What is missing?
Select the best answer.
A city permit office takes 11 weeks to decide building permits, but state law requires 8. It wants design to start now, and you insist on discovery first. Which two questions must discovery answer? (Select two.)
Select exactly 2. 0/2 selected.
You chose a fixed workflow with one Claude summarization step for fault triage over the autonomous agent a vendor demonstrated. A non-technical board asks you to justify it. What should you lead with?
Select the best answer.
A sponsor says a lease-summary assistant must be "always right" before launch. The pilot shows 93 percent field accuracy, and most errors fall in one clause type the review step catches. What should you do?
Select the best answer.
Forty engineers use Claude Code across six repositories with personal instruction files, and some allow any shell command. The platform lead wants shared conventions and safe permissions. What should you set up?
Select the best answer.
Pull requests doubled after a signaling team adopted AI-assisted coding, and two regressions reached production from approved AI-generated changes. Reviewers could not explain either change. The safety manager wants a decision. What fits best?
Select the best answer.
About 3 percent of an assistant's calls to the core banking system fail, mostly when branches open. Each failure returns HTTP 429 within 200 milliseconds. What should you do?
Select the best answer.
A team of 25 engineers will adopt Claude Code this quarter. Several repositories hold system credentials in environment files, and the CIO wants measurable value without a security incident. What should come first?
Select the best answer.
Before you ask
Before you sit the mock
Is this the real CCAR-P exam?
No. Every item here was written for this page from the official CCAR-P Exam Guide and its published domains and objectives. Nothing is recalled from a sitting, nothing is drawn from another provider, and Cred Farmer is not affiliated with Anthropic or Pearson VUE. The items are short scenarios, about the length of the guide's sample questions. Of the 63, 57 are single-answer with four options and 6 ask you to select two of five, saying so in the stem. The questions themselves are ours, and their difficulty is not calibrated against the exam.
How is the mock scored?
You get the raw count of correct answers out of the items you sat, and the same count for each domain beside that domain's official weight. A select-two item counts only when both picks match the key. The exam reports a scaled score with a 720 cut, and no practice percentage converts to it, so this page never shows a scaled score, a pass verdict or a percentile.
Do I need an account to take it?
No. Choose a start mode and the sitting runs in your browser tab. There is no sign-up, no email field and no cookie that follows you. An account only matters if you want the larger signed-in library, and the link for that sits at the bottom of the page rather than in your way.
Is it actually free, or is there a catch?
No catch. Everything on Cred Farmer is free, this page included: no card, no trial, no limit on how many times you sit it. An account costs nothing either; it adds the five timed forms, saved progress and a readiness score capped by your first sitting of each form. The one limit here is inventory: the no-account paper is one paper, so sit it a second time and you will remember some items, and that is when the signed-in forms become the next step.
How many questions are in the mock?
63 in the full sitting, allocated across the 7 official domains in proportion to their weights. The clock runs at the official pace of 63 items in 120 minutes, about 1.9 minutes an item, so the full sitting gets 120 minutes, the same as the exam. The quick sitting draws 30 of them evenly across the domains in 57 minutes, and the untimed review takes the whole paper without a clock.
What happens to my result?
It stays in this tab. Nothing is sent to a server, nothing is stored on your device, and nothing you answer here feeds a readiness score or any aggregate. Close or refresh the page and the result is gone, which is also why it is worth reading the per-option reasons before you leave.
This page is 63 items. The account has 315.
Everything on Cred Farmer is free. An account adds the five timed forms, saved progress and a readiness score: CCAR-P has 5 forms aligned to the current blueprint, 63 items each, 315 in total, plus drills by objective, and your first sitting of each form sets the band the score can reach. No card, no trial clock.
Every item on this page was written for it from the official CCAR-P Exam Guide, not recalled from a sitting and not copied from anyone else's set. Material sold as recalled questions can cost you the credential: read why Claude certification exam dumps are a bad bet. Items were last checked against the guide on . The guide is published on Anthropic's CCAR-P certification page.
Not affiliated with, or endorsed by, Anthropic or Pearson VUE. Exam facts are summarised from publicly published program information and can change. Always confirm against the official exam guide before booking.