Anthropic CCA-F Exam Dumps

Claude Certified Architect Foundations

623 Reviews

Exam Code	CCA-F
Exam Name	Claude Certified Architect Foundations
Questions	300 Questions Answers With Explanation
Update Date	06, 30, 2026
Price	Was : $142.2 Today : $79 Was : $160.2 Today : $89 Was : $178.2 Today : $99

Add To Cart

Demo Questions

Question # 1

A developer implements compaction for a long-running code review session. After compaction, Claudeloses track of files it already reviewed and re-reviews them. What technique preserves the reviewprogress through compaction?

A. Use custom compaction instructions: 'Always preserve the list of reviewed files and their review status'
B. Maintain a scratchpad file (e.g., REVIEW_PROGRESS.md) listing completed reviews, with Claude reading it after compaction to restore state
C. Increase the compaction summary length to capture all reviewed files 
D. Disable compaction and manage context manually 

Show Answer

Question # 2

A startup is designing a customer service agent. They want the agent to detect when a customer isfrustrated based on conversation patterns (multiple repeated questions, escalating language, long waittimes). The detection should trigger a different handling workflow. Where should frustration detectionlogic be implemented?

A. In the system prompt: 'Detect frustration and adjust your approach' 
B. In a PostToolUse hook that monitors the conversation for frustration indicators 
C. As a separate sentiment analysis tool that the agent calls periodically 
D. In the orchestrator as a classifier that analyzes each customer message and adjusts the agent's system prompt or routing based on sentiment signals

Show Answer

Question # 3

A team implements a structured output pipeline where Claude generates quiz questions. Each questionmust have exactly 4 options and one correct answer. They use json_schema strict mode. Occasionally,two options are semantically identical (e.g., 'Use a load balancer' and 'Implement load balancing'). Theschema can't prevent this. What additional quality measure is needed?

A. Add a schema constraint for unique options (uniqueItems in JSON Schema) 
B. Implement a semantic similarity check in the application layer that compares all option pairs and flags near-duplicates for regeneration
C. Use extended thinking to encourage Claude to verify distinctness before generating 
D. Add a system prompt instruction: 'All options must be semantically distinct' 

Show Answer

Question # 4

Your organization is designing a prompt caching strategy. Their application has: (1) a 3K token systemprompt, (2) a 10K token knowledge base, (3) a 2K token conversation history, and (4) user messagesaveraging 200 tokens. They're using Claude Sonnet 4.5 which has a minimum 1024 token cachethreshold. Which portions should be cached?

A. System prompt (3K) + knowledge base (10K) as a single cached prefix, since both are stable and together form a 13K prefix well above the minimum threshold
B. Only the system prompt (3K) since it's stable across all requests 
C. Cache everything including conversation history 
D. Only the knowledge base since it's the largest component 

Show Answer

Question # 5

A developer creates an MCP tool that wraps a SOAP-based legacy API. The SOAP API returns XMLwith deeply nested namespaced elements. What transformation should the MCP tool apply beforereturning results to Claude?

A. Return the raw XML — Claude can parse XML natively 
B. Base64-encode the XML and return it as a string 
C. Convert to a flat key-value pair format 
D. Convert to a simplified JSON structure that strips namespaces and flattens unnecessary nesting, keeping only the business-relevant fields

Show Answer

Question # 6

A senior developer is building a chatbot for a hospital that must NEVER provide specific medicaldiagnoses while still being helpful about general health information. A previous attempt using onlysystem prompt instructions failed when a patient described symptoms in detail and Claude provided atentative diagnosis. What is the correct implementation?

A. Add more emphatic language to the system prompt: 'ABSOLUTELY NEVER provide diagnoses' 
B. Use a two-layer approach: system prompt for general guidance plus a separate classifier model that screens every response before it reaches the patient, blocking anything containing diagnostic statements
C. Implement a PostToolUse classification step that checks every response for diagnostic language and regenerates if detected
D. Limit the conversation length to prevent detailed symptom discussions 

Show Answer

Question # 7

A developer wants Claude Code to output results in a specific format when run in their CI pipeline.They use headless mode with --json-schema. The schema specifies {issues: [{file: string, line: number,severity: string, message: string}]}. However, the output sometimes includes extra explanation textalongside the JSON. What flag ensures only the JSON schema output is returned?

A. The --json-schema flag already guarantees only schema-compliant JSON output in headless mode 
B. --quiet to suppress all logging and only show the result 
C. --json-only to suppress non-JSON output 
D. Pipe the output through jq to extract only the JSON portion 

Show Answer

Question # 8

An architect implements context awareness (formerly citations) to track which sources Claude uses inits responses. The marketing team wants to display source links alongside generated content. Whatdoes the context awareness feature provide?

A. URLs to original source documents that Claude used during training 
B. A list of document IDs that were most relevant to the response 
C. Character-level pointers back to specific locations in the input context, showing exactly which parts of the provided documents Claude referenced for each claim 
D. A confidence score indicating how well-supported each claim is by the input documents 

Show Answer

Question # 9

A developer needs to generate unit tests for a complex function. They have the function implementationand existing tests for similar functions. What prompting strategy produces the highest quality test suite?

A. Send only the function implementation and ask Claude to generate tests 
B. Send the function, its type definitions, and 2-3 examples of test files for similar functions as reference — the examples calibrate Claude's testing style, coverage patterns, and assertion conventions
C. Ask Claude to list all edge cases first, then generate one test per edge case 
D. Send the function and a testing framework documentation page 

Show Answer

Question # 10

A platform engineer is designing an agent that interacts with a third-party API. The API has rate limitsof 100 requests per minute. The agent makes rapid tool calls that can exceed this limit. Where shouldrate limiting be implemented?

A. In the system prompt: 'Wait at least 600ms between API calls' 
B. In a PreToolUse hook that adds a delay between tool invocations 
C. In the MCP server implementation, using a token bucket algorithm that queues requests exceeding the rate limit
D. In Claude's API parameters by setting a request_delay field 

Show Answer

Question # 11

A team builds a legal document analysis agent. The agent needs to cross-reference clauses across a 200-page contract. Their first approach loads the full document and asks Claude to find all cross-references.Results are inconsistent — some cross-references are found, others are missed. What documentprocessing strategy improves reliability?

A. Use a smarter model (Opus instead of Sonnet) for better attention across the full document 
B. Split the document into overlapping sections, have Claude identify cross-references in each section, then merge and deduplicate results across sections
C. Index the document by clause number first, then query specific clause pairs for cross-references 
D. Use extended thinking with a large budget to improve attention throughout the document 

Show Answer

Question # 12

A developer is implementing token counting for their application to monitor costs. They need to counttokens before sending requests to avoid exceeding context limits. Which approach does the Anthropicdocumentation recommend?

A. Use the token counting API endpoint to get exact counts before sending the full request 
B. Use a third-party tokenizer library to estimate token counts locally 
C. Estimate based on character count: ~4 characters per token for English text 
D. Send the request and check the usage field in the response for actual token counts 

Show Answer

Question # 13

An architect wants to implement a 'dry run' mode for their agent where all tool calls are simulatedwithout executing. This is useful for testing agent behavior and estimating costs. Whichimplementation approach is most effective?

A. Add 'this is a dry run, don't actually execute anything' to the system prompt 
B. Replace all tool implementations with mock versions that return realistic simulated responses 
C. Use PreToolUse hooks that intercept all tool calls, log them, and return simulated results without executing the actual tools
D. Set effort='low' to minimize the impact of tool calls during testing 

Show Answer

Question # 14

A developer needs to pass sensitive credentials to an MCP server configured in Claude Code. Theywant the credentials stored securely and not committed to the repository. Where should the credentialsbe configured?

A. In .claude/settings.json with the credentials directly in the MCP server configuration 
B. In environment variables referenced by the MCP server configuration using ${VAR_NAME} expansion syntax 
C. In the CLAUDE.md file with instructions for where to find credentials 
D. Hardcoded in the MCP server source code 

Show Answer

Question # 15

An AI engineer is building a multi-agent system for software testing. Agent A generates test cases,Agent B executes them, and Agent C analyzes results. Agent B discovers a test environmentmisconfiguration that affects all remaining tests. How should Agent B communicate this blocking issueto the orchestrator?

A. Continue executing tests and note failures in the results for Agent C to analyze 
B. Stop execution and return an error message to the orchestrator 
C. Return a structured status with escalation metadata: {status: 'blocked', reason: 'environment_misconfiguration', details: '...', blockedTests: [...], suggestedAction: 'fix_environment_and_retry'}
D. Call Agent A to regenerate test cases that work with the misconfigured environment 

Show Answer

Question # 16

A developer's agent system has a memory of 50 previous customer interactions stored in a database.For each new interaction, they inject all 50 prior interactions into the context for personalization. Thisconsumes 100K tokens, leaving limited room for the current conversation. What is the better approach?

A. Summarize all 50 interactions into a 5K token customer profile that's injected instead 
B. Retrieve only the 3-5 most relevant prior interactions (using embedding similarity or recency) and inject those, keeping context lean while maintaining personalization
C. Increase to a 1M token context model to fit all history 
D. Store all interactions in a tool that Claude can query on demand 

Show Answer

Question # 17

A technical lead is implementing a content moderation system using Claude. They need to classifyuser-generated content as safe, questionable, or unsafe. For legal compliance, they need an audit trailshowing why each decision was made. What API features support this requirement?

A. Enable extended thinking to capture the reasoning, and use structured output for the classification — the thinking blocks provide the audit trail while the structured output provides the consistent decision format
B. Use the Messages API with logging enabled to capture all request/response pairs 
C. Use a separate API call for classification and another for explanation generation 
D. Store the raw API response JSON as the audit trail 

Show Answer

Question # 18

A team uses Claude Code's /init command to generate the initial CLAUDE.md for their project. Aftergeneration, they want to customize it. The /init command has already analyzed the repository. What isthe recommended approach for customization?

A. Delete the generated CLAUDE.md and write one from scratch 
B. Keep the generated CLAUDE.md as a base and iteratively refine it — add project-specific conventions, remove irrelevant sections, and commit the customized version
C. Generate multiple CLAUDE.md files with /init and merge the best parts 
D. Never modify the /init output — it's optimized for the repository 

Show Answer

Question # 19

A developer implements an MCP tool for image processing. The tool accepts an image URL andreturns analysis results. During testing, they find that Claude sometimes passes a data URL (base64-encoded image) instead of an HTTP URL. The tool only supports HTTP URLs. How should the toolhandle this gracefully?

A. Return an error: 'Invalid URL format' without further guidance 
B. Accept data URLs by decoding the base64 content within the tool implementation 
C. Return a structured error with the constraint: {isError: true, message: 'Only HTTP/HTTPS URLs are supported. Data URLs (base64) cannot be processed. Please provide the image's HTTP URL.', isRetryable: true}
D. Silently ignore the data URL and return empty results 

Show Answer

Question # 20

An architect builds an agent that must handle both simple lookups (e.g., 'What is the status of order#1234?') and complex investigations (e.g., 'Why was order #1234 delayed and what is the impact on thecustomer's account?'). Currently both types use the same processing pipeline with extended thinking,making simple lookups slow. What architecture balances speed and depth?

A. Disable extended thinking for all requests and rely on tool results for complex analysis 
B. Let users select 'quick' or 'detailed' mode for their queries 
C. Add a timeout that kills thinking after 3 seconds for simple queries 
D. Use a classifier to route simple lookups (effort='low', no thinking) to a fast path and complex investigations (effort='high', thinking enabled) to a deep-analysis path

Show Answer

Question # 21

An architect needs Claude to generate a complex financial model with nested calculations. They wantto verify that intermediate calculations are correct before using the final result. Extended thinkingwould show the reasoning, but they need the intermediate values in a structured format. What approachcombines both?

A. Enable thinking and ask Claude to include intermediate values in the thinking blocks 
B. Use a structured output schema with nested objects that mirror the calculation hierarchy: {inputs: {...}, intermediate: {step1: {...}, step2: {...}}, final: {...}} with thinking enabled for reasoning visibility
C. Make separate API calls for each calculation step 
D. Use tool calls for each intermediate calculation, collecting results along the way 

Show Answer

Question # 22

A developer is configuring Claude Code settings and notices there are three levels of settingsfiles: .claude/settings.json (project), .claude/settings.local.json (local), and ~/.claude/settings.json(user). They add an MCP server to the project settings. A teammate adds the same server with differentconfiguration to their local settings. Which takes precedence?

A. Project settings (.claude/settings.json) always take precedence 
B. Local settings (.claude/settings.local.json) take precedence over project settings for the same configuration
C. User settings (~/.claude/settings.json) have the highest precedence 
D. The settings are merged with conflicts causing an error 

Show Answer

Question # 23

A solutions engineer is building a RAG (Retrieval Augmented Generation) system that retrievesdocuments and uses Claude to answer questions. After retrieval, they inject 5 relevant document chunkstotaling 15K tokens into the prompt. They want to maximize cache hits across users who ask similarquestions. What prompt structure optimizes caching?

A. System prompt ? document chunks ? user question (documents change per query, breaking the cache)
B. System prompt ? user question ? document chunks 
C. Put all content in a single user message to simplify caching 
D. System prompt (cached) ? common document corpus (cached) ? query-specific chunks ? user question 

Show Answer

Question # 24

A team has their Agent SDK agent configured with max_turns: 10. Their complex workflow typicallyrequires 15-20 tool calls but the agent stops after 10 turns. They need the agent to complete the fullworkflow. What are the two options for handling this?

A. Increase max_turns to 25 and hope it's enough 
B. Split the workflow into exactly 10 steps or fewer 
C. Either increase max_turns, or implement multi-turn looping where the orchestrator checks if the task is complete after max_turns and re-invokes the agent with the current state
D. Remove the max_turns limit entirely to let the agent run indefinitely 

Show Answer

Question # 25

Your team is designing a prompt template for a customer email generator. They want Claude tomaintain the customer's language (English, Spanish, French) automatically without explicit languageinstructions. What is the most effective approach?

A. Detect the customer's language first with a classification call, then include a language instruction 
B. Include a few-shot example in each language to prime Claude for multilingual responses 
C. Use a separate translation API to convert all inputs to English and outputs back to the original language
D. Place the customer's original message prominently in the prompt — Claude naturally mirrors the input language when generating responses

Show Answer