Practice Exam: Claude Certified Architect – Foundations

Question bank: 120 · timed mock exam: 120 min

Official exam page on the vendor's site (opens in a new tab) Official exam guide (opens in a new tab) Official prep course (opens in a new tab)

Exam guide: format, domains and how to prepare

Compare: Claude Certified Architect – Foundations vs Claude Certified Associate – Foundations

  • Agentic Architecture & Orchestration 32 questions
  • Claude Code Configuration & Workflows 24 questions
  • Prompt Engineering & Structured Output 24 questions
  • Tool Design & MCP Integration 22 questions
  • Context Management & Reliability 18 questions
Study mode (free) Exam mode (subscribers)

Free: 25 questions in study mode, all domains, no account.

Subscription: the 120 questions of this certification, the timed exam mode, and all other certifications too.

Subscribe · €9.99 / month

Secure card payment (Stripe) · No commitment · Cancel online

How to prepare for Claude Certified Architect – Foundations with InfraPrep

The 120 questions cover 5 domains: Agentic Architecture & Orchestration, Claude Code Configuration & Workflows, Prompt Engineering & Structured Output, Tool Design & MCP Integration, Context Management & Reliability. 25 of them are free, without an account.

Study mode

Answer at your pace: each answer is corrected straight away, with an explanation and the link to the documentation it relies on.

Exam mode

A timed exam of 120 minutes, with a question map to flag and come back to questions before you submit.

Results by domain

Your score is broken down by domain, so you know what to review first.

Original questions

Every question is written from the public documentation, never copied from the exam. You can contest an answer from study mode.

10 free sample questions

Question 1 Agentic Architecture & Orchestration

A travel company builds a booking assistant directly on the Claude Messages API with four client tools. Its loop ends as soon as a response contains any text block. In testing, the assistant often replies "Let me check the available flights" and the conversation stops there, with no flight search ever executed. How should the loop's control flow be changed?

  1. End the loop after a fixed number of rounds based on the average length of past sessions
  2. End the loop only when Claude's text contains an agreed phrase such as "TASK COMPLETE"
  3. Keep looping while stop_reason is "tool_use", sending back tool_result blocks each round
  4. End the loop when a response contains a text block and no more than one tool_use block
Show the answer and the explanation

Correct answer: C

When Claude wants a client tool, the response carries stop_reason "tool_use" plus tool_use blocks; the application must run the tools, return the results and call the API again. The loop only finishes when Claude produces output with no tool calls.

Source: official documentation

Question 2 Tool Design & MCP Integration

A logistics agent exposes two tools: track_parcel ("Gets parcel info.") and locate_shipment ("Finds a shipment."). Claude often calls the wrong one: track_parcel only accepts carrier tracking numbers, while locate_shipment searches internal shipment IDs. What change does Anthropic's tool-definition guidance recommend first?

  1. Rewrite both descriptions in detail: purpose, when to use or not, inputs, limits
  2. Set tool_choice to auto so that Claude can reconsider which tool to call
  3. Replace the descriptions with input_examples and keep only the tool names
  4. Add a system prompt line telling Claude to prefer locate_shipment in every case
Show the answer and the explanation

Correct answer: A

The guidance is to provide extremely detailed descriptions — what the tool does, when it should and should not be used, what each parameter means and its caveats — aiming for at least 3–4 sentences per tool. Descriptions come first; input_examples are a complement for complex tools.

Source: official documentation

Question 3 Claude Code Configuration & Workflows

A tech lead wrote the team's API conventions in CLAUDE.local.md at the repository root. Claude Code follows them on her machine, but colleagues and the CI job never apply them. The file is listed in .gitignore, as the documentation suggests. What change makes the conventions apply to every clone of the repository?

  1. Import CLAUDE.local.md from ~/.claude/CLAUDE.md using the @path syntax
  2. Run /memory in each colleague's session so that the local file is reloaded
  3. Copy the conventions into ~/.claude/CLAUDE.md on the tech lead's own machine
  4. Move the conventions into ./CLAUDE.md (or ./.claude/CLAUDE.md) and commit it
Show the answer and the explanation

Correct answer: D

A project CLAUDE.md lives in ./CLAUDE.md or ./.claude/CLAUDE.md and is shared with the team through version control. CLAUDE.local.md is meant for private per-project preferences that are not checked in, so it never reaches other clones.

Source: official documentation

Question 4 Prompt Engineering & Structured Output

A company uses Claude Code's managed Code Review on GitHub. Developers complain that it flags too many minor naming remarks in legacy folders, and they start ignoring all its comments. The team wants to state explicitly what should be flagged, at which severity, and how findings are reported — for reviews only, without changing Claude's behavior in other Claude Code tasks. What should they do?

  1. React with a thumbs-down on each unwanted comment so that later reviews adapt
  2. Tell reviewers to reply to each comment asking Claude to be more conservative
  3. Add a REVIEW.md file to the repository describing what to flag and at what severity
  4. Add the review criteria to the root CLAUDE.md so that every task sees them
Show the answer and the explanation

Correct answer: C

REVIEW.md holds review-only instructions; the documentation says to use it to say what the team wants flagged, at what severity, and how findings are reported.

Source: official documentation

Question 5 Context Management & Reliability

In long support conversations, an agent calls get_account_history several times. Each call returns about 60 fields per transaction, of which the agent needs five. After a dozen turns the agent starts overlooking facts the customer stated early on, and costs have tripled. Which change addresses the root cause?

  1. Asking the customer to repeat key facts every few turns
  2. Shaping the tool to return only the fields Claude needs, with filtering or paging
  3. Re-sending the complete system prompt after every tool call
  4. Raising max_tokens so the agent can write longer answers
Show the answer and the explanation

Correct answer: B

Verbose tool results accumulate in the context and crowd out relevant facts. Anthropic recommends that tools include only the fields Claude needs for its next step, and that responses which could use up lots of context use pagination, filtering or truncation.

Source: official documentation

Question 6 Agentic Architecture & Orchestration

A lead research agent delegates to subagents with one-line instructions such as "research heat pump adoption in Europe". Reviewers notice that several subagents run the same searches while some angles of the question are never covered. According to Anthropic's account of its multi-agent research system, what should each delegated task description contain?

  1. The same generic prompt for every subagent, so that their results are comparable
  2. An objective, an output format, guidance on tools and sources, and clear boundaries
  3. The lead agent's complete conversation so far, with no additional instructions
  4. Only a list of search keywords, leaving each subagent free to choose its own goal
Show the answer and the explanation

Correct answer: B

Anthropic states that each subagent needs an objective, an output format, guidance on the tools and sources to use, and clear task boundaries; without detailed task descriptions agents duplicate work and leave gaps.

Source: official documentation

Question 7 Tool Design & MCP Integration

A custom tool that queries an inventory service returns the single word "failed" whenever the upstream service rate-limits it. The agent then either retries instantly in a tight loop or tells the customer the item does not exist. How should the tool result be returned so that Claude can react appropriately?

  1. By throwing an exception that stops the agent loop until an operator restarts it
  2. With is_error set to true and a message saying what went wrong and what to try next
  3. As a raw stack trace from the service client, so that Claude sees full technical detail
  4. As a normal successful result containing an empty list, so the agent can continue
Show the answer and the explanation

Correct answer: B

The tool_result can carry is_error: true, and the documentation asks for instructive error messages that say what went wrong and what Claude should try next (for example, retry after a given delay) instead of a generic "failed".

Source: official documentation

Question 8 Claude Code Configuration & Workflows

A team's /dependency-audit skill reads hundreds of lockfile entries and prints long intermediate listings, which fill the developer's main Claude Code conversation. Which SKILL.md frontmatter setting runs the skill in its own subagent context instead?

  1. context: fork
  2. argument-hint: [package-manager]
  3. disable-model-invocation: true
  4. allowed-tools: Read Grep
Show the answer and the explanation

Correct answer: A

Setting context to fork runs the skill in a forked subagent context. That subagent does not see the conversation history, so the skill's instructions must be self-contained.

Source: official documentation

Question 9 Prompt Engineering & Structured Output

Claude summarizes incident tickets for an on-call dashboard. Despite a long paragraph of formatting instructions, the layout of the summaries keeps drifting from one ticket to the next. According to Anthropic's prompting best practices, which technique is one of the most reliable ways to fix this?

  1. Raising the temperature so the model explores more ways of applying the format
  2. Adding 3–5 relevant, varied example summaries wrapped in tags
  3. Removing the instructions and asking Claude to invent its own consistent format
  4. Repeating the formatting instructions in capital letters at the end of the prompt
Show the answer and the explanation

Correct answer: B

Examples are one of the most reliable ways to steer Claude's output format, tone and structure; the guidance is to include 3–5 relevant, diverse examples, wrapped in tags.

Source: official documentation

Question 10 Context Management & Reliability

A company is piloting a Claude support chatbot and wants a metric that tells it whether the bot hands conversations to people at the right moments. Which measure does Anthropic's customer support guide describe for this?

  1. Escalation efficiency: correct escalations versus ones that should have happened
  2. Average handle time: how long the chatbot takes to close each conversation
  3. Sentiment maintenance: whether sentiment is kept or improved during the chat
  4. Deflection rate: the share of inquiries resolved without any human involvement
Show the answer and the explanation

Correct answer: A

Escalation efficiency measures Claude's ability to recognize when a query needs human intervention, tracked as the share of correctly escalated conversations versus those that should have been escalated but were not.

Source: official documentation

Frequently asked questions

Are there free questions, and do I need an account?

Yes. 25 questions of this certification are free in study mode, across all its domains, without a subscription and without registration. The account is created when you subscribe, at checkout.

How does the subscription work, and can I cancel?

One monthly subscription covers all certifications and the timed practice exams. There is no minimum period. You cancel online, at any time, with the "Cancel your contract" function in My account; cancellation takes effect at the end of the month already paid, and no further payment is taken.

Where do the questions come from?

Every question is original, written from the vendor's public documentation. Each answer comes with an explanation and a link to the documentation page it relies on. InfraPrep is not affiliated with the vendors, and certification names belong to their owners.

In which language are the questions?

The questions are written in English, also on the French version of the site.

Is the catalog updated?

Yes, the catalog changes over time: new certifications and new questions are added.