Claude Certified Architec - Free Practice Quiz

Question 1 of 60

In an agentic loop using the Claude Agent SDK, which field in the API response should be checked to determine whether to continue the loop or terminate it?

A The text content of the assistant response B The number of content blocks in the response C The token count of the response D The stop_reason field in the response object

1 / 60

Question 2 of 60

An agent is configured with 18 available tools. A developer notices that tool selection quality has degraded significantly. What is the best architectural remedy?

A Add more detailed descriptions to all 18 tools B Increase the max_tokens parameter to give Claude more room to reason C Distribute the tools across 4-5 specialized subagents, each with 4-5 focused tools D Enable tool_choice: 'any' to force tool usage on every turn

2 / 60

Question 3 of 60

A customer service agent must enforce a business rule that refunds above $500 require manager approval. Which approach provides the most reliable enforcement?

A Add a clear instruction to the system prompt: 'Never process refunds above $500' B Set a parameter in the tool definition that limits the amount field to 500 C Use few-shot examples in the prompt showing the agent escalating high refunds D Implement a PostToolUse hook that intercepts process_refund calls and blocks amounts over $500

3 / 60

Question 4 of 60

When a subagent in a multi-agent system fails to connect to an external database, what should it return to the coordinator?

A An empty array [] to indicate no results were found B A partial result with a note appended to the text response C A null value so the coordinator can decide how to handle missing data D A structured error object with isError: true, errorCategory, isRetryable fields, and context about what was attempted

4 / 60

Question 5 of 60

A coordinator agent needs to spawn subagents in parallel. Which configuration is required for the coordinator to enable this capability?

A The coordinator must have a higher max_tokens setting than the subagents B The coordinator's allowedTools list must include the 'Task' tool C The coordinator must use claude-sonnet-4-20250514 specifically D Parallel execution requires enabling the 'parallel' flag in the agent configuration

5 / 60

Question 6 of 60

A developer wants to branch a session to explore an alternative implementation approach without polluting the main session context. Which operation should they use?

A --resume to load the previous session state B fork_session to create an isolated branch that diverges from the current context C A new session with --session-name to start fresh D /compact to compress the current context before exploration

6 / 60

Question 7 of 60

Which CLAUDE.md configuration level is the correct location for personal editor preferences that should not be shared with the team?

A Project-level: .claude/CLAUDE.md B Directory-level: src/CLAUDE.md C User-level: ~/.claude/CLAUDE.md D Rules-level: .claude/rules/editor.md

7 / 60

Question 8 of 60

What is the correct way to store API secrets referenced by an MCP server configuration in .mcp.json?

A Hardcode them directly in the .mcp.json env section for simplicity B Store secrets in a separate secrets.json file in the same directory C Base64-encode the secrets before embedding them D Use ${ENV_VAR} environment variable expansion syntax

8 / 60

Question 9 of 60

What is the primary difference between a custom slash command and a skill in Claude Code?

A Commands support MCP servers; skills do not B Skills are stored in .claude/commands/; commands are stored in .claude/skills/ C Commands support YAML frontmatter; skills use plain markdown D Skills can fork context and restrict tool access via allowed-tools; commands run in the current session context

9 / 60

Question 10 of 60

A developer is running Claude Code in a CI/CD pipeline for automated PR review. Which flag is required for non-interactive execution?

A --headless B --no-prompt C -p D --batch

10 / 60

Question 11 of 60

A code review agent that generates code is then asked to review its own output within the same session. What is the primary risk of this pattern?

A It doubles the token cost of the workflow B Review in the same session bypasses the tool call mechanism C The same session cannot access two different file states D The reviewer retains the generator's reasoning context, creating confirmation bias that reduces review effectiveness

11 / 60

Question 12 of 60

Which deployment scenario is best suited for the Message Batches API rather than synchronous API calls?

A A customer-facing chatbot that needs real-time responses B A PR review system that blocks developer workflow while waiting for results C A nightly code audit that processes 10,000 files and can wait up to 24 hours D A real-time fraud detection system that needs sub-second latency

12 / 60

Question 13 of 60

What does the 'lost in the middle' effect mean for Claude context window management?

A Claude forgets information from the beginning of the conversation first B Tool results placed in the middle of a conversation are processed less accurately C Claude's performance degrades when context exceeds 50% capacity D Information placed in the middle of a long context is less likely to be recalled than information at the start or end

13 / 60

Question 14 of 60

A multi-turn customer support session needs to preserve exact customer account details â€” account ID, order number, and specific amounts â€” through a very long conversation. What is the recommended approach?

A Use /compact periodically to compress the conversation while keeping key details B Use progressive summarization to distill the conversation into shorter summaries C Create an immutable 'case facts' block at the start of the context that is never summarized D Repeat the key details at the start of every user message

14 / 60

Question 15 of 60

An automated extraction pipeline tracks overall accuracy at 95%. However, users report that invoice processing is frequently wrong. What is the likely cause and correct diagnostic approach?

A The model has insufficient context; increase max_tokens B The 95% aggregate metric is masking per-document-type failures; stratified metrics by document type are needed C The extraction tool description is too vague; rewrite it with more detail D The validation-retry loop is set to too few retries

15 / 60

Question 16 of 60

When should a customer service agent escalate a conversation to a human agent?

A Whenever the customer expresses negative sentiment B Whenever the model's self-reported confidence drops below 0.7 C When the customer explicitly requests a human, or when a policy gap is detected D After every third interaction to ensure quality

16 / 60

Question 17 of 60

In a multi-agent research system, when delegating a task from a coordinator to a subagent, which context passing approach is correct?

A Pass the full coordinator conversation history so the subagent has complete context B Pass only the specific context relevant to the subagent's assigned subtask C Pass a summary of the coordinator conversation to save tokens D Do not pass any context; let the subagent determine what it needs via tool calls

17 / 60

Question 18 of 60

A developer writes a prompt instructing Claude to 'find all security issues' in a codebase. Production results are inconsistent and developers are experiencing alert fatigue. What is the root cause and correct fix?

A The model needs more examples; add few-shot examples of security issues B The prompt is vague and produces false positives; replace with explicit, measurable criteria like 'identify hardcoded strings matching API key patterns' C The model context window is too small for full codebase analysis; switch to a larger model D The review should use chain-of-thought prompting to improve accuracy

18 / 60

Question 19 of 60

When using tool_use with a forced specific tool to extract structured data from a document, what does this guarantee?

A That the extracted values are semantically correct and match the document B That the output conforms to the defined JSON schema structure C That the extraction will never require a retry D That the model has read and understood the entire document

19 / 60

Question 20 of 60

A validation-retry loop detects that an invoice extraction returned a total field that does not match the sum of line items. What should be appended to the retry prompt?

A A generic instruction: 'Please try again more carefully' B The complete original document for re-extraction from scratch C Specific error details: 'Line items sum ($450) does not match the total field ($500). Please correct the total or the line items.' D A new few-shot example demonstrating correct extraction

20 / 60

Question 21 of 60

A production system uses progressive summarization to manage a long customer support conversation. After multiple summarization passes, the agent references a $99.99 promotional price as $150.00. What caused this failure?

A Progressive summarization compressed and lost the specific promotion details through repeated summarization cycles B The model's context window was exceeded, causing hallucination C The promotional price was never included in the original conversation D The model applied incorrect arithmetic when summarizing

21 / 60

Question 22 of 60

In the CLAUDE.md hierarchy, rules in a directory-level CLAUDE.md file (e.g., src/api/CLAUDE.md) apply to:

A All files in the entire project regardless of location B Only the files immediately in the src/api/ directory, not subdirectories C The src/api/ directory and all directories below it D Only when explicitly imported with @import in the project-level CLAUDE.md

22 / 60

Question 23 of 60

A Claude Code skill is defined with the following YAML frontmatter: 'context: fork' and 'allowed-tools: [Read, Grep]'. What does this configuration guarantee?

A The skill runs synchronously and blocks other operations B The skill shares context with the main session but restricts tool types C The skill creates a new persistent session that can be resumed later D The skill executes in an isolated context that does not affect the main session, and can only use Read and Grep

23 / 60

Question 24 of 60

When tracking data through a multi-agent system where subagents provide conflicting values, which metadata is most important for resolving conflicts?

A The timestamp of when the coordinator started the workflow B The number of tool calls each subagent made C The model version used by each subagent D The source of the data, its confidence level, and when it was retrieved

24 / 60

Question 25 of 60

Which built-in Claude Code tool should be used to find all files matching the pattern '**/*.test.ts' across a codebase?

A Grep B Read C Bash D Glob

25 / 60

Question 26 of 60

A developer needs to search for all usages of the function 'processPayment' across a large TypeScript codebase. Which built-in tool is correct?

A Glob with pattern 'processPayment' B Read on each file sequentially C Grep to search for the pattern 'processPayment' across the source directory D Bash with 'grep -r processPayment src/'

26 / 60

Question 27 of 60

In a CI/CD pipeline, a Claude Code review step needs to output structured JSON that can be parsed by a downstream script. Which flag combination is correct?

A claude -p --output-format json B claude --ci --json C claude --automated --structured D claude --pipeline --parse json

27 / 60

Question 28 of 60

Which tool_choice value forces Claude to use a specific named tool rather than choosing automatically?

A tool_choice: 'any' B tool_choice: 'required' C tool_choice: {'type': 'tool', 'name': 'tool_name'} D tool_choice: 'forced'

28 / 60

Question 29 of 60

A tool named 'search_orders' fails because the order management service is down. The tool returns an empty array []. The agent then tells the customer their order history is empty. What went wrong?

A The tool description was insufficient for the model to understand the error B The max_tokens setting was too low to include the error details C The agent should have used a different tool to check order history D The tool silently suppressed an access failure by returning an empty result, misleading the agent about whether the search was performed

29 / 60

Question 30 of 60

How many few-shot examples is generally recommended as optimal for guiding Claude on ambiguous classification tasks?

A 1 example is sufficient to establish the pattern B 2-4 examples provide the right balance of pattern establishment and context efficiency C 6-8 examples are needed for reliable performance D 10 or more examples are always better for accuracy

30 / 60

Question 31 of 60

A well-designed few-shot prompt for sentiment classification should include which of the following to handle sarcasm correctly?

A At least one example demonstrating an edge case such as sarcasm or mixed sentiment B Only positive and negative examples to keep the prompt concise C Examples with identical output formats even if the inputs differ dramatically D Examples sourced from the actual production data distribution only

31 / 60

Question 32 of 60

A project team wants to share a set of TypeScript coding standards with all developers but also allow individual developers to have personal configuration. What CLAUDE.md structure supports this?

A Use .claude/CLAUDE.md for team TypeScript standards (committed to git) and ~/.claude/CLAUDE.md for personal preferences (not committed) B Put everything in a single .claude/CLAUDE.md shared via git C Store team standards in environment variables accessible to all developers D Create a CLAUDE.md in each developer's home directory with all project rules

32 / 60

Question 33 of 60

During an extended agentic session analyzing a large codebase, the agent begins forgetting earlier constraints and repeating work. What is the recommended remediation strategy?

A Increase max_tokens to give the model more processing capacity B Restart the session completely and re-provide all constraints C Switch to a different model that has a larger context window D Use /compact to compress conversation history and scratchpad files to persist critical state

33 / 60

Question 34 of 60

A subagent is tasked with file-by-file analysis of a large codebase. This task generates extensive tool call output. What architecture keeps the coordinator's context clean?

A Have the coordinator run the analysis directly with a high max_tokens setting B Delegate the verbose exploration to the subagent; the subagent writes findings to scratchpad files; the coordinator reads only the summarized results C Have the coordinator read each file and pass relevant sections to the subagent D Run the analysis as a background process outside the agent context

34 / 60

Question 35 of 60

According to CCA-F exam doctrine, when is it appropriate to use plan mode before executing a task in Claude Code?

A For all tasks, regardless of complexity, to ensure quality B For complex multi-file architectural changes where mistakes are expensive to undo C For simple single-line fixes to ensure the model fully understands the task D Only when the user explicitly requests it

35 / 60

Question 36 of 60

A developer is implementing an iterative refinement workflow using test-driven development with Claude Code. What is the correct first step?

A Write the test first to define expected behavior, then ask Claude to implement to make the test pass B Ask Claude to implement the feature, then write tests afterward C Ask Claude to plan the feature architecture before writing any code D Generate a few-shot prompt with examples of similar features

36 / 60

Question 37 of 60

An MCP server configuration needs Jira credentials for each team member to use their own account. Which .mcp.json configuration is correct?

A Use 'JIRA_TOKEN': '${JIRA_TOKEN}' so each developer's environment variable provides their own token B Set JIRA_TOKEN to the team's shared service account token directly in the file C Create separate .mcp.json files for each developer with hardcoded credentials D Store the token in a .env file referenced by a relative path in .mcp.json

37 / 60

Question 38 of 60

A coordinator agent receives conflicting revenue figures: $1,500,000 from a verified financial database (subagent A) and $1,480,000 extracted from a quarterly PDF (subagent B). How should the conflict be resolved?

A Average the two figures and use $1,490,000 B Prefer the most recent figure based on timestamp alone C Use the value from the higher-confidence source (verified database), log the conflict for audit D Ask the user to manually resolve all data conflicts before continuing

38 / 60

Question 39 of 60

A tool description reads: 'Searches for customers.' An engineer identifies that the model is frequently calling this tool with incorrectly formatted phone numbers. What is the root cause?

A The tool's input_schema is missing required field definitions B The tool description is too vague and lacks input format specifications, examples, and constraint documentation C The model needs more few-shot examples of customer lookup in the main prompt D The agent has too many other tools competing for selection

39 / 60

Question 40 of 60

A structured error response from a tool includes the field 'isRetryable: false' with errorCategory: 'auth'. What should the agent do?

A Retry the tool call immediately with the same parameters B Wait 2 seconds and retry with exponential backoff C Escalate to a human or use an alternative approach â€” do not retry D Log the error and continue with partial results

40 / 60

Question 41 of 60

When implementing a multi-pass code review strategy, why should the cross-file integration pass be a separate step from the per-file local analysis pass?

A To reduce token costs by splitting the analysis across two smaller contexts B Because cross-file issues (broken imports, interface mismatches) cannot be detected when analyzing files in isolation â€” they require knowledge of how files interact C To allow parallel processing of the two analysis phases D Because Claude performs better on shorter prompts

41 / 60

Question 42 of 60

A developer asks: 'Should I use dynamic adaptive decomposition or prompt chaining for a codebase bug analysis task where the number and type of bugs found will determine what steps to take next?' What is the correct approach?

A Prompt chaining, because it provides a predictable structure that is easier to debug B Prompt chaining with a maximum of 10 steps to cap complexity C Both approaches perform identically for this use case D Dynamic adaptive decomposition, because the task has conditional branches where intermediate results determine next steps

42 / 60

Question 43 of 60

In a hub-and-spoke multi-agent architecture, what is the primary purpose of context isolation between subagents?

A To ensure each subagent has only the context relevant to its specific task, preventing confusion from irrelevant information B To improve parallel execution speed C To reduce the total token cost of the workflow D To enable subagents to operate without an internet connection

43 / 60

Question 44 of 60

A SKILL.md file in .claude/skills/analyze/ specifies 'allowed-tools: [Read, Grep]'. During skill execution, the model attempts to call the Bash tool. What happens?

A The Bash call succeeds because allowed-tools is just a suggestion B The Bash call is blocked because allowed-tools restricts the skill's tool access C The skill terminates with an error and cannot be restarted D The Bash call succeeds but is logged as a policy violation

44 / 60

Question 45 of 60

A team wants to organize their Claude Code configuration with separate rule files for TypeScript, testing, and API design standards rather than one large CLAUDE.md. Which mechanism supports this?

A Use @import in the project CLAUDE.md to include modular rule files, and place topic-specific rules in .claude/rules/ B Create separate CLAUDE.md files in each source subdirectory C Combine all rules into a single .claude/rules/all.md file D Use environment variables to switch between rule sets

45 / 60

Question 46 of 60

An engineer discovers that a code review agent frequently dismisses critical findings in the third and fourth review passes of the same session. What architectural change would address this?

A Use separate sessions for each review pass to prevent reasoning context bias from earlier passes B Increase the temperature parameter to reduce repetitive outputs C Add more explicit instructions to not dismiss findings in later passes D Reduce the number of review passes to two maximum

46 / 60

Question 47 of 60

When a customer support agent detects that no existing policy covers a customer's unusual situation, what is the correct escalation response?

A Escalate to a human and provide the full context of what was attempted and why no policy applies B Attempt to apply the most similar policy and proceed with the best guess C Return an error message telling the customer to call back later D Ask the customer to provide more information to find a matching policy

47 / 60

Question 48 of 60

What is the correct stop_reason value that signals Claude wants to invoke a tool and the agentic loop should continue?

A end_turn B stop_sequence C tool_use D max_tokens

48 / 60

Question 49 of 60

After using /compact to compress a long agent session's conversation history, what should the agent do to restore awareness of its current state?

A Restart the entire task from the beginning B Re-read any scratchpad files that contain previously persisted state and findings C Ask the user to summarize what has been accomplished so far D Increase the max_tokens for the next request to compensate

49 / 60

Question 50 of 60

A developer wants team coding standards to apply to all API endpoints in src/api/ but different standards for the rest of the codebase. What is the correct CLAUDE.md structure?

A Create a directory-level CLAUDE.md in src/api/ with the API-specific rules B Add all rules to the project-level .claude/CLAUDE.md with conditional logic C Use environment variables to switch rule sets based on the current working directory D Create separate branches in git for each set of standards

50 / 60

Question 51 of 60

A developer adds a 'document_type' field to an extraction schema with enum values: ['invoice', 'receipt', 'credit_note']. In production, some documents are purchase orders, which are not in the enum. What is the best schema design fix?

A Add 'other' to the enum and add a 'document_type_detail' field for when 'other' is selected B Remove the enum constraint and use a free-form string field C Expand the enum to include every possible document type D Use a numeric code instead of string enum to allow unknown values

51 / 60

Question 52 of 60

Which of the following is an INVALID trigger for escalating a customer support interaction to a human agent?

A The customer explicitly asks to speak to a human representative B No policy in the agent's knowledge base covers the customer's situation C The customer is expressing very negative sentiment about a delayed shipment D The refund amount exceeds the $500 agent authorization limit

52 / 60

Question 53 of 60

When implementing a validation-retry loop with a maximum of 3 retries for structured data extraction, what should the function do after all retries are exhausted?

A Return the best partial result from the last attempt B Return null and log the failure silently C Raise an explicit extraction error that communicates the document could not be processed reliably D Attempt a fourth retry with a different model

53 / 60

Question 54 of 60

A subagent responsible for financial data retrieval returns results without any source, confidence, or timestamp metadata. Why is this a problem for the coordinator?

A The coordinator cannot cache the results without metadata B The coordinator cannot display the results in the UI without formatting metadata C The coordinator's token budget will be exceeded without metadata compression D The coordinator cannot resolve conflicts or assess data reliability when multiple subagents provide overlapping financial data

54 / 60

Question 55 of 60

A production extraction pipeline has a detected_pattern field in its error tracking that shows 80% of retry errors occur on documents where the 'vendor_name' field contains special characters. What should the engineering team do with this insight?

A Use this pattern to trigger targeted human review only for documents matching the pattern, rather than retrying blindly B Increase the max retry count for all documents C Switch to a different extraction model that handles special characters better D Remove the vendor_name field from the required fields list

55 / 60

Question 56 of 60

In the Message Batches API, what is the purpose of the custom_id field on each request?

A It sets the priority level for processing the request within the batch B It associates each individual request with its response for tracking and result retrieval after batch completion C It specifies which model version should process that specific request D It enables deduplication of identical requests within the same batch

56 / 60

Question 57 of 60

A developer wants to use Claude Code to modify a specific function in an existing file without replacing the entire file. Which built-in tool should they use?

A Edit â€” to apply a targeted change to the existing file B Write â€” to save the updated file contents C Bash â€” with a sed command to make the replacement D Read then Write â€” to read the file and write back the modified version

57 / 60

Question 58 of 60

A project's CLAUDE.md uses @import to include multiple rule files. A developer is troubleshooting which rules are active. Where are auto-loaded rule files stored?

A In a .claude/imports/ directory B In the .claude/rules/ directory â€” files here are automatically loaded as rule sets C In ~/.claude/rules/ for user-level auto-loading D In any markdown file with a .rules.md extension anywhere in the project

58 / 60

Question 59 of 60

An agent is analyzing a very large codebase. To keep the coordinator context clean while still getting comprehensive analysis results, which pattern should be used?

A Run the entire analysis in the coordinator session with a high max_tokens cap B Use /compact after every 10 file reads to keep the context size manageable C Have the coordinator read a subset of files and extrapolate findings for the rest D Delegate file-by-file exploration to subagents that write findings to scratchpad files; have the coordinator read only the synthesized summaries

59 / 60

Question 60 of 60

A developer is building a CI/CD code review pipeline where a 'generator' agent writes new authentication code and then a 'reviewer' agent reviews it. The developer considers using --resume to have the reviewer continue the generator's session. Why is this approach wrong?

A The reviewer would retain the generator's reasoning context, creating confirmation bias that undermines the independence of the review B --resume is not available in non-interactive CI mode C --resume creates a new session rather than continuing the existing one D Using --resume doubles the token cost of the combined session

60 / 60

Claude Certified Architect CCA-F Practice Quiz