https://www.passcert.com/CCDV-F.html Page 2 102 questions selected from source version V9.02 CLAUDE CERTIFIED DEVELOPER Question 1 Scenario: Developer Productivity with Claude You are building developer productivity tools using the Claude Agent SDK. The agent helps engineers explore unfamiliar codebases, understand legacy systems, generate boilerplate code, and automate repetitive tasks. It uses built-in tools (Read, Write, Bash, Grep, Glob) and integrates with MCP servers. During a legacy payment module analysis, the coordinator first assigns one subagent to map database tables and another to trace API callers. Both return useful findings. The coordinator then invokes a planning subagent to propose a migration strategy, but the plan ignores the database constraints and caller list already discovered, recommending changes that would break known integrations. What should you change to make this orchestration more reliable? A. Replace the specialized subagents with one long-running generalist subagent that performs discovery and planning together. B. Instruct the planning subagent to infer missing dependencies from file names when prior findings are unavailable. C. Allow subagents to message each other directly so the planning subagent can request missing analysis details. D. Have the coordinator maintain shared investigation state and include relevant prior findings in each subsequent subagent prompt. Answer: D Explanation Coordinator-owned state is central to reliable multi-agent orchestration. In a coordinator-subagent pattern, the coordinator decomposes work, collects results, and decides which findings must be supplied to later subagents so they can produce grounded outputs. The underlying principle is that subagents operate in isolated task contexts. They should not be designed as if they can automatically see prior coordinator conversations or other subagents' outputs. The coordinator should maintain a structured investigation state, such as discovered tables, caller lists, constraints, open questions, and confidence notes, then include the relevant subset in each downstream prompt. Letting subagents communicate directly weakens the hub-and-spoke architecture by hiding information flow from the coordinator. Replacing specialists with one generalist can create context bloat and reduce the value of parallel specialized analysis. Asking a planning subagent to infer missing dependencies from file names is especially risky because it substitutes guesses for evidence. For more on agent orchestration and subagent patterns, see the Agent SDK documentation and the Claude Code Sub-agents documentation. https://www.passcert.com/CCDV-F.html Page 3 CLAUDE CERTIFIED DEVELOPER Question 2 Scenario: Developer Productivity with Claude You are building developer productivity tools using the Claude Agent SDK. The agent helps engineers explore unfamiliar codebases, understand legacy systems, generate boilerplate code, and automate repetitive tasks. It uses built-in tools (Read, Write, Bash, Grep, Glob) and integrates with MCP servers. An engineer asks the agent to explain how the "user export" capability works in a legacy repository. The final answer confidently covers REST controllers and serializers, but misses scheduled exports, admin-triggered jobs, and CLI invocations. Logs show every subagent completed successfully; their prompts were "inspect export controller," "trace export API request," and "summarize export endpoint tests.". What should you change first to improve coverage on similar broad codebase questions? A. Require each subagent to read every file matching export-related terms before returning any findings to the coordinator. B. Run a fixed pipeline that always invokes controller, database, CLI, worker, and test subagents for every codebase query. C. Strengthen the synthesis subagent prompt to infer missing workflows from naming conventions and common framework patterns. D. Revise coordinator planning to identify plausible entry points, then delegate distinct code areas to subagents before synthesis. Answer: D Explanation The root issue is the coordinator's task decomposition, not subagent execution. The subagents completed successfully, but all assigned prompts focused on the REST API path, so the final answer missed other relevant workflows such as scheduled jobs, admin actions, and CLI commands. For broad codebase exploration, the coordinator should first reason about the possible surfaces where a capability may appear, such as controllers, jobs, commands, tests, database models, and integrations. It can then delegate distinct areas to specialized subagents and aggregate their findings into a complete explanation. Asking synthesis to infer missing workflows is an anti-pattern because it encourages unsupported guesses. Having every subagent read every matching file creates context bloat and attention dilution, while a fixed pipeline wastes work and ignores the need for dynamic routing based on query complexity. The underlying principle is that a coordinator-subagent architecture depends on the coordinator to preserve coverage and routing discipline. Learn more about agent orchestration patterns in the Agent SDK documentation. CLAUDE CERTIFIED DEVELOPER Question 3 Scenario: Developer Productivity with Claude You are building developer productivity tools using the Claude Agent SDK. The agent helps engineers explore unfamiliar codebases, understand legacy systems, generate boilerplate code, and automate repetitive tasks. It uses built-in tools (Read, Write, Bash, Grep, Glob) and integrates with MCP servers. An engineer asks the agent to investigate a production stack trace from an unfamiliar legacy service. The trace contains the message "InvalidInventoryTransition" and references a helper named validateTransition, but the repository has hundreds of source files and inconsistent directory naming. What should the agent do first to find the relevant implementation and usage sites efficiently? A. Use Grep to search file contents for the error string and function names, then Read the matching files. B. Use Bash to run ad hoc recursive shell commands, then paste the raw terminal output into context. C. Use Read to load the entire module tree upfront, then ask the agent to infer all relevant references. https://www.passcert.com/CCDV-F.html Page 4 Question 3 continued D. Use Glob to list likely source files by extension, then Read each candidate file until the error appears. Answer: A Explanation Grep is the right first tool when the agent needs to search file contents for known strings such as error messages, function names, import statements, or identifiers. In practice, this lets the agent quickly identify candidate implementation and usage sites, then use Read only on the files that matter. The underlying principle is progressive codebase exploration: start with a narrow content search, then read the smallest useful set of files to understand control flow and dependencies. This preserves context window capacity and avoids attention dilution in large repositories. Glob is useful for matching file paths, such as **/*.test.tsx, but it does not inspect file contents. Loading broad directory trees with Read is an anti-pattern because it adds irrelevant material to context before the agent knows what matters. Using Bash for ad hoc recursive searches can work in some environments, but raw terminal output is often noisy and less controlled than the purpose-built code search tool. For more on Claude Code and built-in development workflows, see Claude Code Overview. CLAUDE CERTIFIED DEVELOPER Question 4 Scenario: Multi-Agent Research System You are building a multi-agent research system using the Claude Agent SDK. A coordinator agent delegates to specialized subagents: one searches the web, one analyzes documents, one synthesizes findings, and one generates reports. The system researches topics and produces comprehensive, cited reports. During a run, the document-analysis subagent asks a source retrieval MCP tool for the full text of a licensed market report. The backend rejects the request because the license permits summaries and short excerpts only. Today the tool returns "access denied," causing the coordinator to retry with paraphrased requests and then mark the source as unavailable. What change would best let the agents handle this case correctly? A. Return a successful empty result set so the synthesis agent can continue without exposing licensing details. B. Return isError true with errorCategory business, isRetryable false, and guidance to request summaries or short excerpts instead. C. Report a permission error and instruct the coordinator to escalate to infrastructure owners for immediate credential repair. D. Classify the rejection as transient and retry with backoff using progressively narrower full-text retrieval queries. Answer: B Explanation Structured MCP errors should give agents enough information to decide whether to retry, revise the request, explain a limitation, or escalate. In this case, the backend is enforcing a licensing constraint: full-text retrieval is not allowed, but summaries and short excerpts are allowed. Returning isError: true with errorCategory: "business", isRetryable: false, and a human-readable explanation lets the coordinator avoid useless retries and choose an allowed retrieval strategy. The key principle is distinguishing business rule violations from transient failures and permission problems. Transient retry logic is appropriate for timeouts or temporary service unavailability, but it cannot overcome a stable contractual limitation. Returning an empty successful result is an anti-pattern because it hides the difference between "no relevant source exists" and "a relevant source exists but cannot be accessed in that form." Treating the case as credential repair also misroutes recovery, since the issue is not missing authorization to perform an otherwise permitted action. Learn more about MCP tool behavior in MCP Tools and Claude tool error handling patterns in Tool Use. https://www.passcert.com/CCDV-F.html Page 5 CLAUDE CERTIFIED DEVELOPER Question 5 Scenario: Customer Support Resolution Agent You are building a customer support resolution agent using the Claude Agent SDK. The agent handles high-ambiguity requests like returns, billing disputes, and account issues. It has access to backend systems through MCP tools (get_customer, lookup_order, process_refund, escalate_to_human). Your target is 80%+ first-contact resolution while knowing when to escalate. A customer starts a chat with: "I don't want to go through a bot. Put me through to a person about order 8421." The message includes enough details that your normal automation would probably resolve the delayed-shipment issue in one or two tool calls. You are tuning the escalation policy and want to avoid both unnecessary deflection and unnecessary automation. What should the agent do next? A. Acknowledge the request, run lookup_order to confirm eligibility, then offer the replacement before considering escalation. B. Use sentiment analysis on the message, escalating only when the customer appears highly frustrated or hostile. C. Ask the customer to explain why human help is needed, then continue autonomously if the reason seems resolvable. D. Call escalate_to_human with the order context and brief summary, without attempting autonomous lookup or replacement processing first. Answer: D Explanation Explicit customer preference is a hard escalation signal. When a customer directly asks for a human agent, the support agent should use escalate_to_human and provide a concise handoff rather than first attempting tool-based resolution. The tradeoff is between maximizing first-contact resolution and maintaining trust. If the customer only expresses frustration but does not request a person, it can be appropriate to acknowledge the frustration and offer help for a straightforward issue. When the customer clearly asks for a human, continuing automation becomes a deflection pattern. Using sentiment as the deciding factor is an anti-pattern because frustration level does not reliably indicate case complexity or customer preference. Asking the customer to justify the request also adds unnecessary friction and can worsen the experience. For implementation guidance on tool-based agent flows and tool result handling, see Tool Use and Agent SDK. CLAUDE CERTIFIED DEVELOPER Question 6 Scenario: Customer Support Resolution Agent You are building a customer support resolution agent using the Claude Agent SDK. The agent handles high-ambiguity requests like returns, billing disputes, and account issues. It has access to backend systems through MCP tools (get_customer, lookup_order, process_refund, escalate_to_human). Your target is 80%+ first-contact resolution while knowing when to escalate. A customer writes: "I was charged twice, my return label never arrived, and I cannot sign in to check the order." The coordinator currently launches billing, returns, and account subagents with the same prompt: "Investigate this customer issue and recommend a resolution." Logs show each subagent calls overlapping tools for the same order, repeats the same facts, and sometimes recommends conflicting next steps. What orchestration change would most effectively improve reliability and efficiency? A. Route every request through returns, billing, and account subagents sequentially, regardless of the customer's stated concerns. B. Send the full customer transcript to every subagent, then use majority agreement to choose the resolution. https://www.passcert.com/CCDV-F.html Page 6 Question 6 continued C. Assign each subagent a distinct issue boundary and shared case facts, then synthesize their non-overlapping findings centrally. D. Allow subagents to message each other directly so they can divide remaining work without coordinator involvement. Answer: C Explanation Correct orchestration in a coordinator-subagent pattern depends on giving subagents complementary assignments rather than duplicative prompts. For a multi-concern support case, the coordinator should pass shared case facts, assign distinct scopes such as billing dispute, return label status, and account access, then aggregate findings into a unified resolution. The underlying tradeoff is between parallel specialization and uncontrolled redundancy. Parallel subagents are valuable when their work is partitioned by issue, subtopic, or source type; they become unreliable when every subagent receives the same broad prompt and independently reaches overlapping conclusions. Using majority agreement over duplicate investigations is an anti-pattern because duplicated context does not create independent coverage. Routing every case through all specialists is also inefficient and fails to scope work. Letting subagents coordinate directly bypasses the coordinator's role in observability, error handling, and controlled information flow. For more on agent orchestration and subagent patterns, see Agent SDK and Claude Code Sub-agents. CLAUDE CERTIFIED DEVELOPER Question 7 Scenario: Customer Support Resolution Agent You are building a customer support resolution agent using the Claude Agent SDK. The agent handles high-ambiguity requests like returns, billing disputes, and account issues. It has access to backend systems through MCP tools (get_customer, lookup_order, process_refund, escalate_to_human). Your target is 80%+ first-contact resolution while knowing when to escalate. During billing disputes, the agent creates condensed case notes before responding or escalating. Human reviewers report that notes sometimes say things like "refund approved" or "customer was promised free return shipping," but they cannot tell whether each claim came from the customer, an order record, a policy document, or the agent's inference. In several escalations, customer allegations were treated as verified backend facts. What change would most effectively reduce this reliability risk? A. Compress each case into a shorter narrative summary and instruct reviewers to ask follow-up questions when provenance seems unclear. B. Store complete raw transcripts and tool outputs only, relying on human reviewers to reconstruct sources when disputes arise. C. Require summaries and handoffs to carry structured claim-source mappings with source type, excerpt or tool field, timestamp, and verification status. D. Ask the agent to include only high-confidence claims in final responses, omitting uncertain details from summaries and escalations. Answer: C Explanation Correct approach: In multi-source support workflows, condensed summaries must preserve claim-source mappings so every material statement remains tied to where it came from. Including the source type, relevant excerpt or tool field, timestamp, and verification status helps distinguish customer assertions, backend records, policy text, and agent conclusions. Underlying principle: Summarization is lossy. When case notes compress a long interaction without structured provenance, source attribution can disappear, and downstream agents or humans may accidentally treat allegations as verified facts. Why the alternatives fail: Shorter narrative summaries intensify provenance loss. Confidence filtering is an anti-pattern https://www.passcert.com/CCDV-F.html Page 7 Question 7 continued because model confidence does not prove factual grounding and can hide uncertainty. Keeping only raw transcripts supports audits, but it forces reviewers to reconstruct evidence manually instead of preserving usable provenance through the workflow. For related implementation patterns, review Agent SDK and Tool Use. CLAUDE CERTIFIED DEVELOPER Question 8 Scenario: Developer Productivity with Claude You are building developer productivity tools using the Claude Agent SDK. The agent helps engineers explore unfamiliar codebases, understand legacy systems, generate boilerplate code, and automate repetitive tasks. It uses built-in tools (Read, Write, Bash, Grep, Glob) and integrates with MCP servers. Your CI job already invokes Claude Code successfully and posts a single pull request summary comment. The team now wants automated inline comments for each finding, but the integration frequently breaks because Claude sometimes adds prose, changes field names, or formats findings as markdown. What change best supports reliable automation? A. Parse Claude's markdown bullets with regular expressions, adding patterns as new review formats appear in CI logs. B. Strengthen the review prompt to demand valid JSON only, then reject responses containing markdown or explanatory text. C. Run Claude Code with --output-format json and --json-schema, then map validated findings to inline PR comments. D. Store the prior summary comment and ask Claude to rewrite it into comment-ready records on a second pass. Answer: C Explanation Reliable CI automation needs structured contracts. When Claude Code output feeds another system, such as an inline pull request comment publisher, the integration should not depend on prose conventions or markdown layout. Using --output-format json with --json-schema gives the pipeline a predictable structure to validate and transform into review comments. Prompt-only formatting is weaker. Asking for JSON in natural language can reduce errors, but it does not provide the same machine-oriented boundary as schema-constrained CLI output. Rejecting responses after the fact shifts the problem into CI failure handling rather than designing the interface correctly. Brittle parsing is an anti-pattern. Regular expressions over markdown bullets or conversational summaries are fragile because the model may change phrasing, ordering, or formatting while still giving a useful review. A second normalization pass can help in some workflows, but it adds cost and latency while remaining less reliable than producing structured output at the source. Learn more about Claude Code automation options in the Claude Code CLI documentation and general structured tool patterns in Tool Use. https://www.passcert.com/CCDV-F.html Page 8 CLAUDE CERTIFIED DEVELOPER Question 9 Scenario: Customer Support Resolution Agent You are building a customer support resolution agent using the Claude Agent SDK. The agent handles high-ambiguity requests like returns, billing disputes, and account issues. It has access to backend systems through MCP tools (get_customer, lookup_order, process_refund, escalate_to_human). Your target is 80%+ first-contact resolution while knowing when to escalate. Production traces show that mixed requests such as "I moved and need to return the charger from order 8831" alternate unpredictably between get_customer and lookup_order as the first call. In some sessions Claude passes an order number to get_customer; in others it passes a customer name to lookup_order. The current tool catalog exposes only terse autogenerated summaries, and no backend errors are occurring. What change should you make first to improve tool selection reliability? A. Update the system prompt to tell Claude to think carefully before selecting tools, without changing schemas or tool metadata. B. Add a keyword router that maps phrases like "order" or "refund" to one tool before Claude sees the request. C. Consolidate get_customer and lookup_order into one broad lookup tool that chooses internally which backend records to retrieve. D. Rewrite each MCP tool description to specify purpose, accepted identifiers, returned fields, edge cases, and boundaries versus related tools. Answer: D Explanation Tool descriptions drive tool selection. When tools have overlapping concepts or accept similar identifier formats, Claude relies heavily on the tool name and description to decide which tool to call. Clear descriptions should explain what the tool does, accepted input formats, returned data, edge cases, and when to use it instead of related tools. Why this works in practice: improving descriptions preserves the existing architecture while giving Claude the information needed to choose reliably between get_customer and lookup_order. This is usually the best first intervention before adding routing layers or redesigning backend interfaces. Why the distractors fail: keyword routing is an anti-pattern because customer support language is ambiguous and multi-intent; vague "think carefully" prompt guidance does not fix missing tool metadata; and consolidating tools into a broad lookup action can blur boundaries instead of clarifying them. The underlying principle is to make tool interfaces explicit and differentiated so model-driven tool choice has the right information available. For more detail, see Anthropic's Tool Use documentation and the MCP Tools documentation. CLAUDE CERTIFIED DEVELOPER Question 10 Scenario: Developer Productivity with Claude You are building developer productivity tools using the Claude Agent SDK. The agent helps engineers explore unfamiliar codebases, understand legacy systems, generate boilerplate code, and automate repetitive tasks. It uses built-in tools (Read, Write, Bash, Grep, Glob) and integrates with MCP servers. Your codebase migration assistant has access to tools for repository metadata extraction, dependency graph analysis, and migration plan generation. Evaluation runs show it sometimes begins dependency analysis or plan generation before collecting canonical package manager and framework versions, producing plans with incorrect assumptions. What change would most reliably enforce the intended first step while preserving model-driven follow-up behavior? A. Inspect the assistant's natural language for phrases indicating metadata completion, then retry when the expected phrase is absent. https://www.passcert.com/CCDV-F.html Page 9 Question 10 continued B. Force the metadata extraction tool with tool_choice for the first request, then continue subsequent analysis in follow-up turns. C. Set tool_choice to any for the first request, relying on tool descriptions to make metadata extraction the likely selection. D. Strengthen the system prompt to say metadata extraction must happen first before any dependency or migration planning tools. Answer: B Explanation Forced tool selection is appropriate when a workflow requires a specific tool to run before other model-selected actions. Using tool_choice: {"type": "tool", "name": "extract_metadata"} or the equivalent named tool form ensures the first request produces the required structured tool call, and the application can then return the tool result before allowing follow-up analysis. The underlying tradeoff is between deterministic sequencing for a required prerequisite and model-driven flexibility for later steps. Once the canonical metadata is in the conversation, subsequent turns can use tool_choice: "auto" or another appropriate setting so Claude can decide whether to analyze dependencies, inspect files, or generate the migration plan. tool_choice: "any" is insufficient because it requires a tool call but does not require the specific metadata tool. Prompt-only instructions remain probabilistic, and natural-language parsing for completion is an anti-pattern because it couples control flow to wording rather than structured tool use. Learn more about tool choice and tool orchestration in Tool Use and broader agent patterns in the Agent SDK. CLAUDE CERTIFIED DEVELOPER Question 11 Scenario: Developer Productivity with Claude You are building developer productivity tools using the Claude Agent SDK. The agent helps engineers explore unfamiliar codebases, understand legacy systems, generate boilerplate code, and automate repetitive tasks. It uses built-in tools (Read, Write, Bash, Grep, Glob) and integrates with MCP servers. An engineer asks the agent to add a caching layer around a legacy authorization module. The codebase has multiple tenants, audit logging, and inconsistent invalidation behavior, and the team has little domain knowledge. Past attempts produced working code that later failed on stale permissions and compliance edge cases. What should the engineer do before asking Claude to modify files? A. Ask Claude to implement the simplest cache wrapper immediately, then iterate only if tests or reviewers expose issues. B. Have Claude first ask targeted questions about invariants, invalidation triggers, audit requirements, and failure modes before proposing implementation. C. Provide a long prose description of desired performance goals and let Claude infer domain rules from surrounding code. D. Tell Claude to copy caching patterns from a different module, assuming consistency across legacy authorization and unrelated services. Answer: B Explanation The best approach is to use the interview pattern before implementation. When engineers are working in an unfamiliar domain, Claude can ask targeted questions to uncover assumptions about invariants, cache invalidation, audit requirements, customer impact, and failure behavior before touching files. The underlying principle is that iterative refinement is not only about fixing code after it exists. It is also about improving the problem definition before implementation, especially when missing requirements could produce correct-looking code with https://www.passcert.com/CCDV-F.html Page 10 Question 11 continued unsafe behavior. Immediate implementation followed by reviewer feedback is an anti-pattern for high-risk legacy behavior because the important constraints may not be visible in the first failing test. Long prose performance goals and copied patterns also fail because they encourage Claude to infer domain rules rather than explicitly validating them. For more on prompting techniques and working effectively with Claude on complex tasks, see Prompt Engineering and Claude Code Overview. CLAUDE CERTIFIED DEVELOPER Question 12 Scenario: Multi-Agent Research System You are building a multi-agent research system using the Claude Agent SDK. A coordinator agent delegates to specialized subagents: one searches the web, one analyzes documents, one synthesizes findings, and one generates reports. The system researches topics and produces comprehensive, cited reports. During production evaluation, reviewers find that first drafts are coherent and cited but often incomplete. For example, a report on AI regulation includes strong EU coverage, weak Asia-Pacific coverage, and unresolved contradictions between two market statistics. The current workflow sends the first synthesis directly to the report generator whenever all subagents return successfully. What change would most effectively improve report completeness without abandoning the coordinator-subagent architecture? A. Have all subagents repeat their original broad assignments twice, then merge duplicate findings before sending results to synthesis. B. Instruct the report generator to hide low-coverage sections unless the synthesis agent marks every subsection as fully supported. C. Raise the coordinator's maximum iteration count to a fixed cap, stopping after five passes regardless of remaining coverage gaps. D. Add a coordinator review loop that checks synthesis coverage, redelegates targeted follow-ups, and reruns synthesis until quality criteria are met. Answer: D Explanation Coordinator-led iterative refinement is the right pattern when a multi-agent research system produces coherent but incomplete outputs. The coordinator should evaluate synthesis quality against explicit coverage criteria, identify gaps or contradictions, redelegate targeted work to search or analysis subagents, and rerun synthesis with the new findings. The underlying architectural principle is that the coordinator owns task decomposition, result aggregation, and quality control. Successful subagent completion only means assigned work finished, not that the overall research question has been adequately covered. Suppressing weak sections in the report generator hides the problem instead of improving evidence quality. Repeating broad assignments creates duplicate work and may still miss the same gaps, while using an arbitrary iteration count as the main stopping rule ignores whether coverage is actually sufficient. Learn more about agent orchestration patterns in the Agent SDK documentation. https://www.passcert.com/CCDV-F.html Page 11 CLAUDE CERTIFIED DEVELOPER Question 13 Scenario: Code Generation with Claude Code You are using Claude Code to accelerate software development. Your team uses it for code generation, refactoring, debugging, and documentation. You need to integrate it into your development workflow with custom slash commands, CLAUDE.md configurations, and understand when to use plan mode vs direct execution. Your team added a custom review workflow that summarizes generated code changes into structured records containing modified files, test coverage, security impact, rollback risk, and reviewer recommendation. Each record currently has one overall confidence score. In pilot runs, high-scoring records often have accurate file lists and test summaries but occasionally miss rollback or security concerns, which are the highest-risk fields. What change would best support safe routing of low-risk records while preserving reviewer attention for uncertain areas? A. Reduce confidence variance by forcing Claude Code to output only high, medium, or low confidence for each record. B. Automatically approve records above a fixed overall confidence threshold after adding instructions to be conservative on risky changes. C. Send every generated record to senior reviewers until the workflow reaches a stable overall accuracy percentage for one month. D. Require confidence per critical record field, then calibrate routing thresholds using a labeled set of previously reviewed changes. Answer: D Explanation Field-level calibration is the right pattern when different parts of a generated artifact have different risk and accuracy profiles. In this situation, accurate file lists and test summaries do not compensate for missed rollback or security concerns, so routing decisions need confidence signals tied to each critical field rather than one aggregate score. The underlying reliability principle is that confidence must be validated, not merely reported. A labeled set of previously reviewed changes lets the team measure how well each confidence score predicts correctness, then set review thresholds that preserve human attention for uncertain or high-risk fields. Using a fixed overall threshold is an anti-pattern because it can automate the exact failures that reviewers are finding. Reviewing everything avoids automation risk but does not solve reviewer capacity allocation, and aggregate accuracy can hide weak segments. Coarse labels such as high, medium, and low are also insufficient unless they are calibrated against labeled outcomes. For related practices, see Claude Code Overview and Prompt Engineering. https://www.passcert.com/CCDV-F.html Page 12 CLAUDE CERTIFIED DEVELOPER Question 14 Scenario: Customer Support Resolution Agent You are building a customer support resolution agent using the Claude Agent SDK. The agent handles high-ambiguity requests like returns, billing disputes, and account issues. It has access to backend systems through MCP tools (get_customer, lookup_order, process_refund, escalate_to_human). Your target is 80%+ first-contact resolution while knowing when to escalate. Production logs show inconsistent tool selection for billing-related messages. Routine refund requests are sometimes escalated, while requests requiring policy review are sometimes sent directly to process_refund. There are no backend errors, and the tool registry shows broad descriptions with little detail about side effects, eligible inputs, or boundaries between refund handling and human escalation. What is the best first change to improve reliability? A. Add a keyword router that sends messages containing refund to process_refund and messages containing supervisor to escalate_to_human. B. Rewrite the affected tool descriptions to specify purposes, required inputs, side effects, examples, and boundaries between refund and escalation workflows. C. Merge process_refund and escalate_to_human into one handle_billing_case tool that internally decides the correct backend action. D. Raise temperature slightly so the agent explores alternative tool choices instead of repeatedly choosing the same billing action. Answer: B Explanation Clear tool descriptions are the main interface Claude uses to decide which MCP tool is appropriate for a given request. In this situation, the problem is not backend failure or unavailable tooling, it is unreliable selection among tools whose boundaries are underspecified. Descriptions should explain when to use each tool, required inputs, outputs, side effects, examples, and when a similar tool is more appropriate. The key architectural principle is that tool interfaces are part of the agent design, not just developer documentation. For tools with business impact, such as process_refund and escalate_to_human, descriptions must make consequences and eligibility criteria explicit so the model can reason correctly before acting. Keyword routing is a brittle anti-pattern because customer support language is ambiguous and context-dependent. Merging tools into a generic handler can obscure important boundaries, while changing temperature only changes randomness and can reduce reliability. For more guidance, see Tool Use and MCP Tools. CLAUDE CERTIFIED DEVELOPER Question 15 Scenario: Code Generation with Claude Code You are using Claude Code to accelerate software development. Your team uses it for code generation, refactoring, debugging, and documentation. You need to integrate it into your development workflow with custom slash commands, CLAUDE.md configurations, and understand when to use plan mode vs direct execution. A three-month pilot dashboard reports 96% reviewer acceptance for Claude Code generated artifacts. Review notes show most accepted outputs are README edits and changelogs, while rejected outputs cluster around OpenAPI response schemas and database migration rollback notes. Leadership wants to stop manual review for any output where Claude reports confidence above 0.9. What is the most reliable next step? A. Adopt the 0.9 confidence threshold globally, then monitor incidents and re-enable review if production defects increase. B. Require reviewers to check only low-confidence outputs, since high-confidence Claude Code results https://www.passcert.com/CCDV-F.html Page 13 Question 15 continued already passed aggregate evaluation. C. Increase the overall random sample size for two more sprints, then decide using the combined acceptance rate. D. Break down validation results by artifact category and generated field, then reduce review only for consistently reliable segments. Answer: D Explanation Segmented validation is the reliable way to decide where human review can be safely reduced. A high aggregate acceptance rate can mask poor performance on specific artifact types, such as API schemas, or specific generated fields, such as rollback instructions. Confidence calibration requires comparing model-reported confidence against labeled reviewer outcomes, then setting review thresholds only where the data shows consistent reliability. In practice, teams should analyze accuracy by artifact category and field, then keep human review for segments with elevated error rates or ambiguous requirements. The anti-patterns are using a global confidence cutoff, relying on a larger pooled acceptance rate, or waiting for production incidents before restoring review. These approaches ignore the observed clustering of failures and can undermine trust in otherwise useful automation. Learn more about Claude Code workflows in Claude Code Overview and structured evaluation practices in Prompt Engineering. CLAUDE CERTIFIED DEVELOPER Question 16 Scenario: Claude Code for Continuous Integration You are integrating Claude Code into your CI/CD pipeline. The system runs automated code reviews, generates test cases, and provides feedback on pull requests. You need to design prompts that provide actionable feedback and minimize false positives. A CI review job recently added strict accessibility guidance intended only for React UI files. After deployment, Claude starts flagging backend route handlers and migration scripts for missing ARIA labels because the guidance is loaded on every review. You need the guidance to apply automatically only when relevant changed files are in scope. What should you change? A. Add a root CLAUDE.md instruction telling Claude to ignore accessibility guidance unless filenames look like React components. B. Create a .claude/commands/accessibility.md command and require CI to invoke it only for pull requests containing UI files. C. Use @import from the project CLAUDE.md and ask Claude to decide whether the imported accessibility file applies. D. Move the accessibility guidance into a .claude/rules/file with YAML frontmatter paths matching React component globs. Answer: D Explanation Path-specific rules are the right fit when guidance should load only for files matching particular path patterns. In Claude Code, rule files in .claude/rules/can use YAML frontmatter with a paths field containing glob patterns, such as paths: ["src/**/*.tsx", "components/**/*.tsx"], so conventions activate only for matching files. The underlying tradeoff is precision versus always-loaded context. Keeping specialized guidance in a root CLAUDE.md or importing it unconditionally can pollute reviews for unrelated files, increasing token usage and false positives. A slash command can support an explicit workflow, but it is not the mechanism for automatic path-based convention loading. https://www.passcert.com/CCDV-F.html Page 14 Question 16 continued Prompting Claude to inf