Claude Certified Architect Foundations

CCAF Exam Questions & Answers

15 / 153 questions

Free CCAF Exam Questions & Answers · Detailed ExplanationsPage 1 of 2

Customer Support Resolution AgentAgentic Architecture & Orchestration

A support service is investigating whether a late-delivery fee should be reversed. Claude returns a short explanation and a lookup_order tool_use block with ID tu_41; stop_reason is tool_use. The lookup succeeds, returning delivery timestamps but no refund decision. The customer has already authorized this read-only investigation. The integration must let Claude assess the result and continue without another customer message. Which continuation correctly preserves the tool conversation?

Select one answerARun the lookup, append its output as an ordinary user message, and ask Claude to finish the investigation.BAppend the complete assistant response, return a tool_result linked to tu_41, and request Claude's next response.CRun the lookup, replace the assistant response with a delivery-status summary, and resume from the shortened conversation.DRun the lookup and return its delivery timestamps as the completed response, retaining the tool request in the case log.

Customer Support Resolution AgentTool Design & MCP Integration

A refund tool reports the same error text for a temporary service outage and an order that is ineligible under policy. The agent retries both, wasting time on ineligible orders. Under the tool-result contract, execution failures use isError. Which TWO output changes support appropriate recovery and communication?

Select two answersAReturn a structured error category and an explicit retryability indicator for the failure.BInclude a human-readable explanation and relevant context for the permitted next action.CEncode recovery instructions only in HTTP status names exposed as raw text.DReturn success with an empty refund record for policy rejection.

Customer Support Resolution AgentContext Management & Reliability

A support conversation covers two disputed orders. After three compactions, the summary says a customer is owed about $60 soon. Earlier verified tool results recorded a $38.40 delivery charge on order O17 and a separate $21.95 duplicate charge on O42; the agent promised an update on 14 October, not a refund by that date. Both issues remain open. The context budget cannot retain every raw tool response, but later actions must use the exact amounts, issue statuses and commitment. Which context design best meets those requirements?

Select one answerAKeep the latest narrative summary and retrieve the full transcript only after an action fails a backend validation check.BInclude the last two tool responses verbatim and instruct each new summary to preserve any other important details.CStore one combined disputed total with the promised date, and reconstruct each order's details when a refund is selected.DMaintain updated per-issue case facts and the exact commitment outside the compacted narrative, injecting them into subsequent prompts.

Code Generation with Claude CodeClaude Code Configuration & Workflows

A team's root CLAUDE.md interleaves API, testing and release conventions. All three sets apply throughout the repository and must reach every developer through version control. The team wants separately owned topic files, a short root file, and automatic loading of new topic files without maintaining an import list or asking developers to invoke a command. Which organization best satisfies these requirements?

Select one answerAPlace the shared topic files in .claude/rules/ without paths filters, keeping only brief universal context in root CLAUDE.md.BPlace the shared topic files in docs/standards/ and add an @import entry to root CLAUDE.md for each file.CTurn each topic into a project skill and ask developers to invoke the relevant skill before starting every task.DInstall the topic files under each developer's ~/.claude/rules/ and synchronize those directories through team onboarding.

Code Generation with Claude CodeClaude Code Configuration & Workflows

A project documentation skill inspects many source files and writes a draft. Its exploration should run outside the main conversation, returning only a concise result. Shell commands must be unavailable during this skill's execution, but developers must retain Bash for other tasks. The installed Claude Code build supports skill-level disallowed-tools; its general-purpose agent can read and write files. Which TWO skill frontmatter changes directly meet the isolation and tool-availability requirements?

Select two answersASet allowed-tools to Read, Grep, Glob and Write while leaving the skill's execution context unchanged.BSet disallowed-tools to Bash for the skill invocation.CSet context to fork and agent to general-purpose for the skill task.DAdd a Bash deny rule to the project permission settings used by every task.

Multi-Agent Research SystemAgentic Architecture & Orchestration

A research coordinator must compare retention, deletion and subprocessor policies for three vendors in both EU and US deployments. Its proposed worker assignments cover each vendor's global retention documentation, while the report template already contains all six vendor-region sections. Each worker has adequate tools and a query budget, but no research has started. The coordinator must detect missing investigations before spending that budget. Which planning artifact most directly exposes whether the assignments can support the requested comparison?

Select one answerAA report outline specifying the six sections and the citation format required in each section.BA resource estimate listing each worker's query budget and expected completion time.CA source-quality checklist requiring every worker to use recent vendor documentation and return citations.DA coverage matrix mapping each vendor-region-policy requirement to an assigned investigation and its required evidence.

Multi-Agent Research SystemContext Management & Reliability

A research pipeline extracts cited measurements, normalizes their units, then combines them into regional comparisons before a report agent writes the narrative. The archive and bibliography survive, but comparisons no longer record which measurements produced them. An auditor must trace each final comparison to the contributing evidence, including its document location and measurement date. The report agent receives the final handoff and source catalog, not the original documents. Which TWO changes preserve the needed attribution through these transformations?

Select two answersAGive extracted measurements stable evidence identifiers linked to their source location, supporting excerpt and measurement date.BAttach the complete bibliography for each region to its report section and retain the original documents in the archive.CCarry one source reference and date on each regional comparison, using the most recent contributing measurement as representative evidence.DRequire every normalization and synthesis output to carry links from each derived comparison to its contributing evidence records.

Developer Productivity with ClaudeTool Design & MCP Integration

Before changing a rounding rule, a developer assistant must identify consumers of calculateTotal in billing/math.ts. Grep finds its definition and a barrel file that exports it as invoiceAmount. Reading that barrel also reveals a wrapper exported as priceOrder; the wrapper calls the same calculation. Searches for calculateTotal find no application call sites. The assistant has Grep, Glob and Read and wants a targeted investigation rather than reading the entire repository. What should it do next?

Select one answerARepeat the calculateTotal content search with more surrounding lines, then treat files in those matches as the consumer list.BUse Glob to collect files with billing in their paths, then inspect those files for calls to the calculation.CTrace the re-exports and wrapper with Read, search for each outward-facing name with Grep, and inspect the resulting call sites.DInspect files that directly import billing/math.ts and use that direct-import set as the complete affected consumer list.

Developer Productivity with ClaudeTool Design & MCP Integration

An assistant explores a codebase through an MCP endpoint returning a numeric relevance score and an unlabeled snippet. It cannot tell whether a hit is a definition, caller or test, so its dependency explanations are unreliable. What output contract best supports the investigation?

Select one answerAReturn only a more precise relevance score.BReturn all matching files in full without indicating why they matched.CReturn classified hits with symbol, location and supporting evidence.DHave the model infer the role from the endpoint name.

Developer Productivity with ClaudeAgentic Architecture & Orchestration

A developer assistant has one afternoon to improve regression tests for an unfamiliar order service. Initial exploration shows that the endpoint with the most incidents delegates to a shared retry helper. That helper is also used by invoice creation, where a duplicated operation has different consequences. Existing endpoint tests mock the helper entirely. The assistant has not yet traced the retry state or the invoice consumer. Which decomposition best uses the remaining investigation and test-writing time?

Select one answerAComplete the original endpoint-only test plan, keeping the helper mocked so the investigation stays within its first boundary.BTrace the helper and its consequential consumers, prioritize observable failure cases, and revise the test tasks as those dependencies become clear.CAssign equal test-generation time to every file touched by the endpoint, then merge the resulting unit tests.DGenerate tests from the endpoint signature first, then ask a reviewer to find missing interactions after the tests are complete.

Claude Code for Continuous IntegrationClaude Code Configuration & Workflows

A CI reviewer receives a pull-request diff through standard input. Its read permissions are already approved, and it must complete one review without a terminal user. The comment publisher expects a single JSON response envelope and reads a schema-conforming findings array from structured_output; it cannot consume a stream or extract JSON from prose. The installed CLI supports --json-schema. Which invocation design matches this existing publisher contract?

Select one answerAUse claude -p with --output-format json and describe the findings schema only in the review prompt.BUse claude -p with --output-format stream-json and pass the findings schema through --json-schema.CUse claude -p with --output-format json and pass the findings schema through --json-schema.DUse claude -p with text output, require JSON in the prompt, and send standard output directly to the publisher.

Claude Code for Continuous IntegrationPrompt Engineering & Structured Output

Claude generates a patch that caches authorization decisions by account ID. In the generation conversation it assumes every user in an account has the same permissions, although the product requirements describe user-specific roles. A same-session review repeatedly accepts that assumption. The requirements, relevant source files and focused tests are available as separate artifacts. The team wants a reviewer that can challenge the assumption without losing the information needed to judge the patch. Which review setup best addresses this failure?

Select one answerAStart an independent review context with the requirements, patch, related code and tests, excluding the generation rationale.BContinue the generation conversation, add a stricter review checklist and require an explicit approval decision.CStart a new review context containing only the changed lines, withholding the requirements to avoid influencing the reviewer.DStart a new review context with the artifacts and the generator's full explanation so the reviewer can preserve its intended design.

Claude Code for Continuous IntegrationPrompt Engineering & Structured Output

A large PR changes validation and serialization across several modules. Full-PR reviews spend most attention on the first files, while the final consumer receives a changed field type. Local details and cross-module consistency both matter. Which review structure most directly addresses this attention problem?

Select one answerARepeat the broad review and publish only findings every run repeats.BCombine focused local passes with producer-consumer integration review.CChoose a larger context window and leave the attention structure unchanged.DReview only the final consumer because it is where failures appear.

Structured Data ExtractionPrompt Engineering & Structured Output

An invoice pipeline flags records whose extracted line items do not sum to the extracted total. Human investigation has found two causes: the model sometimes copies the balance-due figure into total, and some invoices really contain inconsistent printed totals. The source files are retained, but the review queue currently shows only extracted numbers and a failed-validation flag. Reviewers need to distinguish those causes before a correction changes the accounting record. Which review payload best supports that decision?

Select one answerARe-extract until the arithmetic check passes, sending the first consistent record and its retry count to the review queue.BKeep the extracted source-stated total and independently calculated total separate, with the discrepancy and source excerpts or locators for the relevant fields.CReplace the extracted total with the line-item sum, retaining the original extracted total in an audit field without its source location.DKeep both totals and attach the model's confidence score, routing lower-confidence records for source inspection first.

Structured Data ExtractionContext Management & Reliability

An extraction service reviews all fields below model confidence 0.90 and automatically accepts the rest. Labeled checks of accepted fields show 99.4% accuracy for standard invoice totals but 93.0% for renewal dates in narrative contracts. The pooled accepted accuracy is 99.1% because invoices dominate the workload. Management requires evidence that each automated field-document segment meets its acceptance target before review volume is reduced. What is the most defensible next step?

Select one answerAEvaluate confidence against labeled outcomes within each field-document segment, then choose and validate segment-specific automation thresholds.BRaise the global threshold until pooled accepted accuracy reaches the target, then apply it to every document type.CRetain 0.90 for every segment and expand only the review of fields below that score, since accepted fields already passed the threshold.DAsk the model to use stricter confidence scoring for contracts, then reduce review when fewer contract fields exceed 0.90.

Complete your exam preparation

Unlock the full CCAF question bank

Practise free with answers, explanations, and Community discussions. Join Community to unlock the complete question bank and timed mock exams.

Join Community Get full CCAF access

Claude Certified Architect certification guide

Claude Certified Architect – Foundations is Anthropic’s technical certification for solution architects building production applications with Claude. ExamCertPro provides independent preparation for the Claude certification exam.

Claude Certified Architect exam topics

Our Foundations preparation covers agentic architecture and orchestration, tool design and MCP integration, Claude Code configuration and workflows, prompt engineering and structured output, and context management and reliability. Check the official certification page for the current Claude architect certification requirements.

Claude certification guide CCAF exam questions CCAF Mock Exams Free exam questions and answers

What is the CCAF exam?

CCAF stands for Claude Certified Architect – Foundations.

Official resources: Anthropic · Claude Certified Architect – Foundations certification · Anthropic certification announcement