The CCAR-P uses question shapes most practice courses never show you: short
one-line scenarios, yes/no matrices where you judge five statements
independently, and matching items where the same option can be used more than
once. Quiz engines can't render the last two, so they live here — free,
untimed, every answer explained.
A lender runs 400 applications a day through the same five checks, in the same order, every time. Which pattern fits?
2 / 16multiple choiceD1 · d1.4
A site-acquisition review needs environmental, zoning, and structural analysis, each with its own data sources and specialist context, merged into one recommendation. Which design?
3 / 16multiple choiceD2 · d2.3
An application puts the current timestamp at the top of its system prompt, followed by 9,000 static tokens. Prompt caching is enabled but never hits. Why?
4 / 16multiple choiceD2 · d2.1
A hard rule buried mid-prompt in 12,000 tokens of product context is followed inconsistently, and rewording it has not helped. What most improves adherence?
5 / 16multiple choiceD3 · d3.1
Twelve internal systems must be reachable by several different Claude applications, and each system's team wants to own its own integration. Which approach fits?
6 / 16multiple choiceD3 · d3.4
A contract assistant retrieves fixed 300-token chunks. Clauses come back without the definitions they depend on, and answer quality suffers. What is the best change?
7 / 16multiple choiceD4 · d4.2
A team needs an evaluation set for a support assistant before launch. Which composition gives the most trustworthy signal?
8 / 16multiple choiceD4 · d4.5
A revised prompt beats the current one by six points on the offline eval suite. What should happen before it replaces production?
9 / 16multiple choiceD5 · d5.3
An agent drafts emails, updates internal records, and can issue purchase orders up to $50,000. Where does the approval gate belong?
10 / 16multiple choiceD6 · d6.1
A client opens an engagement with "we need a chatbot for our intranet." What should the architect do first?
11 / 16yes / noD1 · d1.3
For each characteristic, decide whether it favours an autonomous agent over a fixed workflow.
The processing steps are the same on every request and known in advance.
The path forward depends on what each intermediate result reveals.
Every step must be individually auditable and reproducible.
The work benefits from the model choosing its own sequence of tool calls.
Inputs always arrive in one fixed format against one fixed schema.
12 / 16yes / noD5 · d5.1
A RAG agent is being hardened against prompt injection carried inside retrieved documents. For each measure, decide whether it reduces the risk.
Delimit retrieved content as untrusted data and instruct the model never to treat it as instructions.
Move to a larger model, since more capable models cannot be injected.
Restrict the agent's tool permissions so a successful injection cannot exfiltrate data or destroy state.
Set temperature to zero so the model ignores adversarial content.
Log every retrieved document so injections can be reviewed after the fact.
13 / 16yes / noD6 · d6.2
You are closing the discovery phase of an AI engagement. For each item, decide whether it should be documented before design work begins.
Agreed, measurable success criteria tied to the business problem.
The final production system prompt.
Data availability, quality, access constraints, and compliance obligations.
The specific model version that will be pinned in production.
The visual design of the user interface.
14 / 16matchD1 · d1.3
For each scenario, identify the most appropriate architectural pattern.
Draft a personalised renewal letter for each policyholder, pulling their coverage record in for context.
Onboard a new vendor: the same six verification steps, in the same order, every time.
Track down why a nightly batch job started failing, where each query determines what to check next.
Prepare a site-acquisition package needing environmental, zoning, and structural review, each with its own data sources, merged into one recommendation.
Convert a scanned invoice into structured fields using a supplied field dictionary.
Options may be used more than once.
15 / 16matchD3 · d3.1
For each integration need, identify the most suitable connection mechanism.
Three separate Claude products each need to read and write the same CRM contact records.
A scheduled reconciliation script pushes settled balances into the ledger with no model in the path.
A hospital's scheduling agent must negotiate transport slots with an ambulance operator's own agent, neither exposing internal systems.
A newly built pricing-rules service should be usable by any agent the company builds this year or next.
An existing OCR pipeline calls the model once per page to extract totals, inside a fixed internal flow.
Options may be used more than once.
16 / 16matchD5 · d5.1
For each risk, identify the most appropriate primary control.
The assistant must never emit a customer's full payment card number, under any circumstance.
Each recommended dosage change carries clinical consequences and must be signed off individually.
Compliance must reconstruct, a year later, exactly what the agent did on a disputed order.
Generated product copy must never state a warranty term the company does not offer, and must be stopped before publication.
The team wants to know within a day if answer quality degrades after a model version change.
Options may be used more than once.
Ready for full-length practice? Form 1 is free on the
practice page — 63 questions,
timed or untimed. The complete set of six tests (332 questions) is
on Udemy.