Claude Certified Architect — Foundations: 60-Question Interactive Practice Exam

Claude Certified Architect — Foundations: 60-Question Interactive Practice Exam

The Claude Certified Architect — Foundations exam is not simply a vocabulary test.

It is a decision-making exam.

You may know what a subagent is. You may understand MCP. You may have memorized the purpose of CLAUDE.md, Plan Mode, structured outputs, hooks, tools, retries, and context windows.

But the real question is this:

Can you choose the right architecture when several answers appear technically possible?

That is where many candidates lose points.

This 60-question practice exam is designed to test the five major areas emphasized in the Claude Certified Architect — Foundations blueprint:

1. Agentic Architecture and Orchestration

2. Claude Code Configuration and Workflows

3. Prompt Engineering and Structured Output

4. Tool Design and MCP Integration

5. Context Management and Reliability

The exam is scenario-based because the real certification is likely to reward production judgment, not simple recall.

This is an independent practice resource and is not an official Anthropic exam.

How to Use This Practice Exam

Copy this entire article into Claude or ChatGPT.

Then add the following instruction:

Act as my Claude Certified Architect — Foundations exam instructor. Ask me one question at a time, beginning with Question 1. Do not reveal the answer until I respond. After each response, tell me whether I am correct, explain the reasoning, update my score, and then present the next question. Track my performance by exam domain. When all 60 questions are complete, give me my final score, identify my weakest domains, and create a focused study plan.

You can also ask the AI to deliver the exam in blocks:

Give me Questions 1 through 15 first. Grade each answer individually before continuing.

For a more difficult version, use:

Do not give me hints. Make me explain my reasoning before revealing the correct answer.

For a rapid review version, use:

Ask me each question, accept only A, B, C, or D, and keep the explanations brief.

The source cram sheet emphasizes the agentic loop, programmatic enforcement, Claude Code configuration, structured outputs, MCP tool routing, escalation, ambiguity resolution, and production-grade decision-making.

Domain 1: Agentic Architecture and Orchestration

Question 1 — Completing the Agentic Loop

A customer-support agent calls an order lookup tool. The tool returns the requested order information successfully, and the application immediately displays the raw tool result to the customer.

What is the most important architectural correction?

A. Add more detail to the tool description

B. Return the tool result to Claude so it can continue reasoning and determine the next action

C. Convert the tool response into XML

D. Create a separate formatting agent for every tool response

Question 2 — Mandatory Business Controls

A company requires identity verification before any refund over $500. The system prompt instructs the agent to verify identity, but the agent occasionally skips the step during long conversations.

What is the best solution?

A. Repeat the instruction several times in the prompt

B. Add more few-shot examples

C. Place a deterministic authorization gate before the refund tool

D. Increase the model’s reasoning budget

Question 3 — Agent or Workflow?

A nightly process downloads a known report, validates it against a fixed schema, stores valid records, and sends invalid records to a review queue.

The process and rules are stable.

What is the best architecture?

A. A fully autonomous multi-agent system

B. A deterministic workflow that uses Claude only where unstructured interpretation is needed

C. A peer-to-peer agent network

D. A conversational assistant with persistent memory

Question 4 — Parallel Research

A research application must investigate five independent subjects and produce one consolidated report with consistent citations.

Which architecture is most appropriate?

A. A coordinator that assigns work to parallel research subagents and synthesizes their findings

B. A sequential pipeline in which each researcher rewrites the previous output

C. Five agents communicating freely without a coordinator

D. One formatting agent with access to five search tools

Question 5 — Over-Decomposition

A team creates separate code-review agents for naming, comments, imports, semicolons, blank lines, logging, function length, file length, and error messages.

The system is expensive, slow, and produces conflicting recommendations.

What is the best redesign?

A. Give every agent more context

B. Add another agent to resolve disagreements

C. Consolidate related checks and use deterministic linters where possible

D. Run every agent twice and use majority voting

Question 6 — Error Classification

An agent selects the correct database tool and supplies valid arguments. The request fails because the database connection pool is exhausted.

How should the failure primarily be classified?

A. Reasoning error

B. Prompt error

C. Environment or infrastructure error

D. Context-management error

Question 7 — Safe Retry Design

A read-only payment-status tool times out. Temporary network failures are common.

What is the best first response?

A. Retry forever

B. Retry a limited number of times with backoff, then return a structured failure or escalate

C. Ask Claude to infer the payment status

D. Permanently disable the tool

Question 8 — High-Risk Ambiguity

A medical-support assistant receives a description that could indicate either a routine issue or an emergency.

What should the architecture prioritize?

A. Continue questioning until the model reaches complete confidence

B. Choose the statistically most likely diagnosis

C. Trigger a safety-aware escalation based on uncertainty and predefined risk criteria

D. Let several agents vote on a diagnosis

Question 9 — Partial Failure

A coordinator launches four research agents. Three succeed, while one fails because its data source is unavailable.

What is the strongest production behavior?

A. Discard the entire result

B. Silently omit the failed topic

C. Preserve successful findings, retry or use an approved fallback, and disclose the unresolved limitation

D. Ask another agent to invent the missing information

Question 10 — Context Passing

A coordinator sends every subagent the entire 200-page conversation, even though each subagent needs only one small section.

What is the primary concern?

A. Subagents cannot receive coordinator context

B. Excessive context increases cost, noise, and loss of focus

C. Every subagent must receive identical context

D. Context prevents parallel execution

Question 11 — Choosing a Pipeline

A document-processing system must perform these dependent steps:

1. Extract text

2. Classify the document

3. Apply rules based on the classification

4. Generate a final summary

Which orchestration pattern fits best?

A. Pipeline

B. Peer-to-peer

C. Uncoordinated multi-agent group

D. Random tool selection

Question 12 — Coordinator Responsibility

A coordinator delegates research to multiple subagents.

Which responsibility should remain with the coordinator?

A. Performing every detailed search itself

B. Combining results, resolving conflicts, and ensuring the final response satisfies the original request

C. Duplicating every subagent’s work

D. Removing all provenance before returning the report

Question 13 — Reasoning Failure

An agent has access to the correct tool, but chooses an unrelated tool because it misunderstands the user’s request.

What type of failure is this?

A. Environment error

B. Reasoning or routing error

C. Transport error

D. Database-capacity error

Question 14 — Human-in-the-Loop

A legal-document agent is uncertain whether a clause creates a major financial obligation.

What is the best design?

A. Force the agent to make a final decision

B. Allow it to omit the clause

C. Escalate the uncertain, high-impact decision to an authorized human reviewer

D. Increase temperature to encourage creativity

Question 15 — When Agents Are Overkill

A service must rename uploaded files according to a fixed pattern and reject unsupported extensions.

What is the best approach?

A. Multi-agent orchestration

B. Deterministic application logic

C. A research subagent

D. A conversational workflow with long-term memory

Domain 2: Claude Code Configuration and Workflows

Question 16 — Plan Mode

A developer asks Claude Code to replace authentication across twelve services while preserving backward compatibility.

What should happen first?

A. Modify the first authentication file found

B. Use Plan Mode to inspect dependencies and propose a sequence before editing

C. Create a slash command called /authentication

D. Clear all repository context

Question 17 — Direct Execution

A developer asks Claude Code to fix a misspelled label in one known React component.

What is the best approach?

A. Start a multi-agent architecture review

B. Enter an extended planning session

C. Make the scoped change and run the relevant validation

D. Rewrite the repository instructions

Question 18 — Durable Repository Guidance

A team wants Claude Code to consistently follow its testing commands, naming standards, architectural boundaries, and pull-request expectations.

Where should this guidance primarily live?

A. In temporary chat messages

B. In version-controlled Claude instruction files with appropriate scoping

C. Inside every source file

D. Only in CI logs

Question 19 — Skill or Command?

A team has a detailed security-review methodology with reference material and evaluation criteria. Claude should recognize when it is relevant and load it automatically.

What is the best fit?

A. A skill

B. A basic shell alias

C. An MCP transport

D. A temporary chat response

Question 20 — Custom Slash Command

A team repeatedly performs the same explicit task: generate a release-readiness checklist when a developer types a named command.

What is the best fit?

A. A custom slash command

B. A database resource

C. A peer-to-peer agent topology

D. A context compaction rule

Question 21 — Subagent Boundary

A task requires an isolated security review with its own instructions, limited tools, and separate context.

What is the best fit?

A. A subagent

B. A spelling rule

C. A hook that always exits successfully

D. A single comment in README.md

Question 22 — Hooks

A team must block Claude Code from writing to protected deployment files unless a validation script succeeds.

What should enforce this?

A. A polite prompt reminder

B. A lifecycle hook or programmatic gate

C. A longer system prompt

D. A few-shot example

Question 23 — Instruction Scope

A repository contains special rules that apply only to database migration files.

What is the best configuration strategy?

A. Put the rules in every prompt

B. Use path-specific or scoped repository instructions

C. Put them only in a developer’s personal notes

D. Create a separate repository

Question 24 — Ignored Content

A repository contains generated files and large build artifacts that Claude Code should not inspect.

What should the team use?

A. A repository ignore mechanism such as .claudeignore

B. More subagents

C. A higher temperature

D. A research coordinator

Question 25 — CI/CD Execution

A CI pipeline must invoke Claude Code non-interactively and collect machine-readable output.

Which configuration is most appropriate?

A. Interactive chat mode with plain text output

B. Non-interactive execution with the -p flag and JSON output

C. A shared local conversation

D. Manual copy and paste

Question 26 — Session Isolation

Several pull requests are analyzed concurrently. Output from one run occasionally appears influenced by another run.

What is the most important correction?

A. Use one persistent shared session

B. Isolate every CI execution and provide only the context required for that run

C. Increase prompt size

D. Ask the model to ignore earlier information

Question 27 — Team and Personal Instructions

A developer has personal preferences, while the repository has mandatory team standards.

Which guidance should control shared project behavior?

A. Personal preferences should always override the repository

B. Version-controlled project and team instructions should govern shared behavior

C. The longest instruction should win

D. The newest chat message should permanently change the repository

Question 28 — Large Ambiguous Change

A user says, “Modernize the entire application.”

What should Claude Code do first?

A. Begin editing random files

B. Clarify intent, inspect the repository, define scope, and develop a plan

C. Delete deprecated-looking code

D. Update dependencies without review

Question 29 — Validation After Editing

Claude Code completes a scoped bug fix.

What should happen next?

A. Stop immediately after changing the file

B. Run the most relevant tests, checks, or validation commands

C. Create three subagents to reread the prompt

D. Clear all context before reviewing the result

Question 30 — Reusable Capability

A team wants to distribute the same repository-health assessment methodology across many projects.

What is the strongest approach?

A. Re-enter the full prompt manually in every repository

B. Package it as a reusable skill or installable shared capability

C. Place it in one developer’s chat history

D. Store it only in screenshots

Domain 3: Prompt Engineering and Structured Output

Question 31 — Schema Enforcement

An extraction system must always return a customer name, account number, and nullable cancellation date in valid JSON.

What is the best technique?

A. Ask for “clean output”

B. Use an explicit JSON schema with required fields and nullable types

C. Ask the model to use its best judgment

D. Parse a prose paragraph with regular expressions

Question 32 — Missing Data

A document does not contain a cancellation date.

What should the extractor return?

A. A guessed date

B. An empty but valid nullable value according to the schema

C. A date from a similar document

D. A prose explanation instead of JSON

Question 33 — Few-Shot Examples

When are few-shot examples most helpful?

A. When they demonstrate the exact classification or output pattern expected

B. When they replace all validation

C. When they are unrelated but stylistically interesting

D. When the task is already fully deterministic and requires no model

Question 34 — Self-Correction

A model occasionally returns malformed JSON.

What is the strongest recovery pattern?

A. Accept malformed output

B. Detect schema failure and run a constrained correction pass

C. Increase creativity

D. Remove the schema

Question 35 — Vague Review Criteria

A prompt says, “Review this code carefully.”

Why is this weaker than explicit review criteria?

A. The model cannot read code

B. “Carefully” does not define concrete pass-and-fail conditions

C. Explicit criteria always reduce accuracy

D. Vague prompts are required for production systems

Question 36 — Categorical Evaluation

A compliance reviewer must check whether a document includes consent language, a retention period, an owner, and an escalation path.

What prompt design is strongest?

A. “Tell me whether this looks good”

B. A checklist with explicit categories and evidence requirements

C. “Be thoughtful”

D. “Use your intuition”

Question 37 — Batch Processing

A system must extract structured fields from 10,000 similar documents.

What design is strongest?

A. Process everything in one enormous prompt

B. Batch the work, validate each result, and track failures independently

C. Skip validation to reduce latency

D. Use one conversation session forever

Question 38 — Multi-Pass Review

A high-value report must be factually accurate, clearly written, and policy compliant.

What is the strongest design?

A. Ask for all goals in one vague sentence

B. Use separate passes or reviewers for factuality, clarity, and compliance

C. Increase temperature

D. Remove the source material

Question 39 — Hallucination Prevention

An extraction system must not invent fields that are absent from the source.

What should the prompt and schema require?

A. Guess plausible values

B. Return null, an empty permitted field, or an explicit not-found status

C. Use external facts without attribution

D. Fill every field regardless of evidence

Question 40 — Output Contract

A downstream application expects a fixed enum for urgency: low, medium, or high.

What is the strongest prompt design?

A. Ask for any descriptive urgency phrase

B. Define the enum explicitly in the output schema

C. Convert all outputs manually

D. Ask for a paragraph explaining urgency

Question 41 — Conflicting Instructions

A prompt contains both “return JSON only” and “explain your reasoning in detail.”

What is the main problem?

A. The instructions conflict with the output contract

B. JSON cannot contain strings

C. Reasoning always improves machine parsing

D. Models ignore all format instructions

Question 42 — Evidence-Based Classification

A model classifies customer complaints into policy categories.

What should improve reliability?

A. Require the category and supporting evidence from the source text

B. Ask only for a category name

C. Increase response length

D. Remove examples

Domain 4: Tool Design and MCP Integration

Question 43 — Primary Routing Signal

Claude has access to several similar tools.

What is the primary signal used to determine when a tool should be selected?

A. The tool’s description

B. The shortest function name

C. The system prompt alone

D. The order in which tools were registered

Question 44 — Strong Tool Description

Which tool description is best?

A. “Searches stuff”

B. “Useful search function”

C. “Searches active customer orders by order ID or customer email; use for order status, shipping, and fulfillment questions; do not use for invoices or refunds”

D. “Tool number three”

Question 45 — Similar Tools

An MCP server provides both search_orders and search_invoices.

How should routing confusion be reduced?

A. Give both tools identical descriptions

B. Clearly define inputs, examples, and boundaries between the tools

C. Rename both tools search

D. Hide all descriptions

Question 46 — Structured Tool Errors

A tool fails because a required record does not exist.

What should it return?

A. A random empty string

B. A structured error containing the error type, message, and recoverability information

C. A fabricated record

D. An unhandled exception only

Question 47 — MCP Primitives

Which set contains the core MCP primitives?

A. Tools, resources, and prompts

B. Threads, databases, and stylesheets

C. Hooks, branches, and commits

D. Models, invoices, and queues

Question 48 — MCP Resource

A server exposes a company policy document for Claude to read as context.

Which MCP primitive best fits?

A. Resource

B. Tool

C. Retry policy

D. Transport error

Question 49 — MCP Tool

A server exposes an operation that creates a support ticket.

Which MCP primitive best fits?

A. Resource

B. Tool

C. Static prompt only

D. Context cache

Question 50 — Destructive Tool Safety

A tool can permanently delete customer data.

What is the strongest design?

A. Use a vague description

B. Require explicit authorization, validate inputs, and add programmatic safeguards

C. Let the model decide without confirmation

D. Retry deletion automatically after every failure

Question 51 — Tool Input Design

A tool requires an account ID and date range.

What is the strongest interface?

A. One unstructured text field with no validation

B. Typed, clearly named parameters with documented formats and constraints

C. Hidden input assumptions

D. A function with no description

Question 52 — Tool Output Design

A downstream agent must distinguish successful results, empty results, validation errors, and temporary failures.

What should the tool return?

A. One plain-text sentence for every condition

B. A consistent structured response with status and typed fields

C. Nothing on failure

D. A guessed success response

Question 53 — Tool Boundary

A tool description says it can “handle all company operations.”

What is the main problem?

A. The scope is too broad and provides poor routing guidance

B. Broad descriptions always improve reliability

C. Tool descriptions do not matter

D. The function name will solve the ambiguity

Question 54 — Advanced MCP Behavior

A server needs to inform clients that a long-running resource has changed.

Which advanced MCP capability is most relevant?

A. Notifications

B. CSS

C. SQL joins

D. Git rebasing

Domain 5: Context Management and Reliability

Question 55 — Long Conversation

A long-running session becomes noisy, expensive, and less accurate.

What is the best response?

A. Continue adding all historical content forever

B. Reduce, summarize, segment, or selectively preserve the context

C. Increase temperature

D. Duplicate the entire history

Question 56 — Retrieval

An assistant must answer questions from thousands of internal documents.

What is the best general approach?

A. Insert every document into every prompt

B. Retrieve the most relevant content and provide it with the request

C. Ask the model to rely on memory

D. Randomly sample documents

Question 57 — Prompt Caching

A large stable instruction block is reused across many requests.

What is a likely benefit of prompt caching?

A. Reduced repeated processing cost and latency

B. Guaranteed factual accuracy

C. Elimination of all context limits

D. Automatic tool authorization

Question 58 — Confidence Calibration

An assistant has weak evidence for a factual claim.

What is the best behavior?

A. State the claim with complete certainty

B. Communicate uncertainty and request or retrieve stronger evidence

C. Invent a supporting citation

D. Hide the uncertainty

Question 59 — Provenance

A research system synthesizes information from several sources.

What should it preserve?

A. Only the final prose

B. The relationship between claims and their supporting sources

C. The model’s guesses

D. Unattributed copied text

Question 60 — Final Production Decision

Two proposed solutions can complete the task.

Solution A is faster to prototype but depends entirely on prompt instructions for mandatory controls.

Solution B uses explicit validation, structured outputs, scoped context, bounded retries, and programmatic enforcement for critical rules.

Which answer best reflects the mindset of the certification exam?

A. Choose Solution A because it uses fewer components

B. Choose Solution B because it is more reliable, observable, and production-ready

C. Choose whichever response is longer

D. Choose the solution with the most agents

Final Exam Instructions for Claude or ChatGPT

After pasting the exam, use this prompt:

Begin with Question 1. Ask me only one question at a time. Wait for my answer before grading it. After each answer:

1. Tell me whether I am correct

2. Explain why the correct answer is best

3. Explain why my answer was weaker if I was wrong

4. Update my total score

5. Track my score by domain

6. Present the next question

Do not reveal future answers. At the end, give me:

• My overall percentage

• My score by domain

• The concepts I repeatedly misunderstood

• Ten targeted remediation questions

• A final readiness rating: Not Ready, Developing, Nearly Ready, or Exam Ready

The Mindset That Matters

The strongest answer will not always be the answer that uses the most AI.

Sometimes the right answer is an agent.

Sometimes it is a coordinator with subagents.

Sometimes it is a structured extraction loop.

Sometimes it is a deterministic workflow.

Sometimes it is simply a validation gate that the model cannot bypass.

That distinction matters.

The certification is likely to reward architects who understand where Claude adds judgment, where software must provide enforcement, and how the two work together in a production system.

Do not study only definitions.

Study decisions.

Do not ask only, “Can this work?”

Ask:

What fails under load?

What becomes expensive?

What happens when the tool returns an error?

What happens when the model is uncertain?

What must never be left to prompt compliance alone?

That is the difference between understanding Claude and architecting with Claude.

:::

The article can also be converted into a cleaner Wix format with callout boxes, section dividers, a scoring table, and a downloadable answer-sheet layout.