Claude Certification Guide CCAR-F

Practice questions

105 scenario questions across all five domains, mixed by default so you have to recognise which domain a question belongs to — half the difficulty on the real exam. Every answer explains why the wrong options are wrong, which is where the reasoning actually lives.

0 / 0correct
0current streak
0best streak

Browse all 105 questions

Each question has its own page with the answer and a full explanation.

Agentic Architecture & Orchestration 27
  1. A team is writing the control flow for an agentic loop that calls Claude with a set of tools. Which loop…
  2. In a Claude agentic loop, which signal should the loop treat as meaning the model has finished its work?
  3. A support agent built with the Agent SDK calls lookup_order with the same order ID three or four times in a…
  4. A research system produces a report on "the economic effects of an aging population." Every subagent returns…
  5. A coordinator sends every query through web search, document analysis, synthesis, and report generation…
  6. Reviewers find that about a third of research reports have thinly supported sections, even though the…
  7. A team proposes letting the web search subagent send its results directly to the synthesis subagent instead…
  8. A coordinator needs results from three independent search subagents, each covering a different subtopic. Logs…
  9. A synthesis subagent is invoked with the prompt "Write a cited summary of what the search and document…
  10. Final reports from a research system contain accurate claims, but citations often point to the wrong document…
  11. A developer configures AgentDefinitions for search and synthesis subagents and writes a coordinator prompt…
  12. A coordinator instructs its document analysis subagent: "Open the first PDF, read pages 1–5, extract three…
  13. A support agent's system prompt requires calling get_customer to verify identity before process_refund. After…
  14. When the support agent escalates, it calls escalate_to_human with the note "Customer needs help with refund…
  15. A customer writes: "My order arrived damaged, I was billed twice for it, and I need to update my shipping…
  16. A support agent's system prompt contains the four rules below. Which two should be enforced programmatically…
  17. Where should enforcement of a rule like "never transfer more than $10,000 without a second approval"…
  18. A support agent pulls order data from an MCP tool that returns Unix timestamps and numeric status codes, and…
  19. Policy requires human approval for refunds over $500. A team adds a hook that rejects process_refund calls…
  20. An agent team is deciding where to use Agent SDK hooks instead of prompt instructions. Which two requirements…
  21. A CI step reviews pull requests by sending all changed files to Claude in one prompt. On a 25-file PR, the…
  22. An engineer asks an agent to "add comprehensive tests to this legacy codebase," which has 300 modules, almost…
  23. Which two tasks are better suited to dynamic adaptive decomposition than to a fixed prompt chain? Select two.
  24. A developer spends a day investigating a memory leak with Claude Code and wants to continue that same…
  25. After an agent has analyzed a payments service, the team wants to compare two refactoring approaches…
  26. Two days into an investigation session, a large refactor is merged that moves and rewrites most of the files…
  27. A developer resumes a named session in which the agent had analyzed an authentication module. Overnight, they…
Claude Code Configuration & Workflows 22
  1. A monorepo has one root CLAUDE.md that three package teams keep editing to add build steps that apply only…
  2. A 1,200-line root CLAUDE.md mixes testing, API, deployment, and frontend standards. Package maintainers say…
  3. A new engineer reports that Claude Code ignores the team's API naming conventions in their sessions. The tech…
  4. For the same developer on the same branch, Claude follows the repository's logging convention in some…
  5. Your project CLAUDE.md holds the team's naming conventions plus a 15-step release procedure that is used…
  6. A team's /analyze-deps skill reads dozens of files and prints a detailed report on each one. Developers run…
  7. A team's /scaffold-tests skill should only create test files and must never run shell commands. Developers…
  8. A senior engineer wants the team's shared /review skill to apply her own extra security checklist when she…
  9. The root CLAUDE.md contains 250 lines of Terraform conventions. Most sessions are spent on application code…
  10. Each of a platform's 14 services keeps its SQL migrations in its own db/migrations/ directory, for example…
  11. Storybook story files (*.stories.tsx) sit next to their components in about 60 directories. The team pasted…
  12. A production error's stack trace points to a missing null check on line 88 of invoice_formatter.py. The fix…
  13. During a multi-phase task (find every use of a deprecated config flag, design a replacement, then implement…
  14. You used plan mode to investigate migrating from one ORM to another across 48 files. The plan lists the…
  15. You asked Claude to convert legacy log lines into structured JSON events. Your prose description is detailed…
  16. A Claude-written migration script mostly works, but it writes the string "null" when a source column is null…
  17. A developer with no caching experience asks Claude to put a cache in front of a slow pricing service. They're…
  18. Review of a Claude-generated payment module finds three problems: retries re-send requests that already…
  19. Your CI job invokes Claude Code to review pull requests. The job sometimes hangs until it times out, and when…
  20. Your CI review runs on every push to a pull request. Developers complain that each push reposts the same…
  21. In your pipeline, the Claude Code session that implements a fix is then asked to review its own changes…
  22. A CI job uses Claude Code to generate unit tests for changed modules. Reviewers report that about half of the…
Prompt Engineering & Structured Output 20
  1. A CI review bot is told to "check that code comments are accurate." Developers dismiss most of its comment…
  2. A review bot's security findings are almost always accepted, but its "performance" findings are dismissed…
  3. A review prompt defines severity as "critical: severe problems; major: significant problems; minor: small…
  4. A review prompt contains a detailed paragraph specifying that every finding must include location, issue…
  5. A test-generation reviewer is inconsistent about whether an untested error branch counts as a coverage gap…
  6. An extractor pulls sample size from research papers. It is accurate when the paper has a methodology section…
  7. A pipeline prompts Claude to "reply only with a JSON object," but roughly 5% of responses include a leading…
  8. A ticket classifier uses a closed enum of five categories. New kinds of tickets, such as accessibility…
  9. An intake pipeline has three extraction tools (invoice, receipt, contract) and doesn't know in advance which…
  10. A pipeline must run extract_metadata on each document before any enrichment tool, because enrichment depends…
  11. An invoice extraction pipeline retries failed validations by sending the document, the failed extraction, and…
  12. A vendor-onboarding extractor retries up to five times when tax_id is missing. These retries almost never…
  13. An invoice extractor uses tool use with a strict schema, and every response passes schema validation…
  14. Developers dismiss about 40% of an automated reviewer's findings. The team tracks the dismissal rate but…
  15. Documents arrive throughout the day and must be processed within 30 hours of arrival. The pipeline uses the…
  16. A batch of 5,000 contract extractions completes with 120 failures. Most of the failed contracts were long…
  17. A team wants to move workloads to the Message Batches API for its 50% cost savings. Which workloads are good…
  18. A Claude session generates a database migration, then in the same session is asked to "carefully review the…
  19. A team moved its large pull request reviews to one pass per file, and review depth is now consistent. Then a…
  20. A review pipeline posts every finding directly to pull requests. The team wants to send uncertain findings to…
Tool Design & MCP Integration 18
  1. In a research system, the analysis subagent has two tools: analyze_content, described as "Analyzes the…
  2. A support agent has a tool whose entire description is "Queries the v2 inventory microservice via internal…
  3. A developer productivity agent has a web_search tool and an MCP code_search tool. Both have clear, distinct…
  4. A support agent's lookup_order tool wraps its database call in a catch-all that returns an empty list on any…
  5. An MCP tool calling an external catalog API sometimes gets "item not found" and sometimes gets a timeout…
  6. A process_refund tool rejects refunds above a $500 policy limit. The agent currently retries the call three…
  7. In a research system, a document-fetch subagent sometimes gets HTTP 503 responses from a publisher's site…
  8. Every subagent in a multi-agent research system was given the same 18 tools. Tool selection errors are…
  9. A document analysis subagent has a generic fetch_url tool. It sometimes loads search result pages and blog…
  10. An extraction pipeline must call extract_metadata before any enrichment tools, because they need its output…
  11. A classification step offers the model three tools and must always produce one of them as a structured tool…
  12. A team's .mcp.json, committed to the repository, configures a GitHub MCP server with an engineer's personal…
  13. A team is standardizing Claude Code MCP setup. They need Jira for everyone, and one engineer wants to try an…
  14. An agent is connected to an internal MCP server whose search_code tool indexes all 40 service repositories…
  15. A support agent connected to an internal knowledge-base MCP server spends many tool calls at the start of…
  16. Asked to explain how a checkout request is processed in an unfamiliar 600-file service, an agent starts by…
  17. An agent needs to change one timeout value in a YAML file that has three identical service blocks, each…
  18. Before deprecating formatCurrency, an agent Greps for that name and finds 6 callers. A teammate points out…
Context Management & Reliability 18
  1. A returns agent calls lookup_order several times per session. Each call appends the tool's full 40+ field…
  2. In long support sessions where customers raise several issues, older turns are progressively summarized. Late…
  3. A coordinator concatenates twelve subagent reports into one synthesis prompt, in the order they arrive…
  4. A customer's first message is: "I want to talk to a real person about my bill." The agent currently looks up…
  5. A customer asks the agent to apply the loyalty-program discount retroactively to an order placed two weeks…
  6. A lookup_customer call by name returns three matching accounts. The agent picks the account with the most…
  7. A compliance-research subagent queries a regulatory filings database. When the database times out, the…
  8. You are defining how a web-search subagent handles failures so the coordinator can make good recovery…
  9. A research coordinator continues with partial results when two of seven sources remain unavailable after…
  10. Two hours into a Claude Code session spent reading files in a large codebase, Claude now describes "the…
  11. An agent must assess the impact of changing refund logic in a large monorepo. Running every search and file…
  12. A multi-agent exploration of a monorepo crashes about 70% of the way through. Restarting it repeats every…
  13. An invoice-extraction pipeline shows 97% field accuracy overall, and the team plans to remove human review…
  14. An extraction model outputs a confidence score for each field, and fields below 0.8 go to human review. An…
  15. A team has automated acceptance of high-confidence extractions. With a small review team, they want to keep…
  16. Reviewers cannot trace most statistics in a research system's final reports back to a source. Subagents…
  17. A document-analysis subagent finds two credible studies reporting different adoption rates, 34% and 52%…
  18. A synthesis agent reports a "major contradiction" between a survey finding 22% adoption and another finding…