60 questions in 120 minutes, set inside 4 production scenarios drawn at random from 6, about 15 questions each. Most questions have one best answer and three plausible distractors; some ask you to select TWO. Scored out of 1000, with 720 widely reported as the pass mark. Five domains, led by Agentic Architecture & Orchestration at 27%.
The domains and their weights
Agentic Architecture & Orchestration 27%
- Run agent loops on
stop_reason: "tool_use" means run the tool and continue, "end_turn" means finished. Never parse text or rely on iteration caps. - Hub-and-spoke: subagents only talk to the coordinator, and they inherit nothing. Put every fact they need in their prompt.
- Spawn independent subagents in parallel by emitting several Task calls in one turn.
- High-stakes rules (money, security, compliance) need programmatic gates or hooks, not prompt instructions.
- Trace failures upstream: narrow research output usually means the coordinator decomposed the task too narrowly.
Practise 32 Agentic Architecture & Orchestration questions →
Tool Design & MCP Integration 18%
- Tool descriptions are THE mechanism for tool selection. Fix vague descriptions before reaching for classifiers or few-shot examples.
- Return structured errors: transient, validation, business or permission, with an isRetryable flag.
- An empty result is a valid answer, not a failure. Do not retry it.
- Keep each agent to roughly 4-5 role-scoped tools. Use tool_choice "any" or a named tool to force a call.
- Project MCP servers live in
.mcp.json(shared, use ${ENV_VARS} for secrets); personal ones in~/.claude.json.
Practise 22 Tool Design & MCP Integration questions →
Claude Code Configuration & Workflows 20%
- Team standards belong in the project CLAUDE.md. User-level
~/.claude/CLAUDE.mdis not shared, which is a classic exam trap. .claude/rules/files withpaths:globs apply conventions to matching files anywhere in the repo.- Skills are on-demand workflows;
context: forkkeeps noisy output out of the main conversation. - Plan mode for large, multi-file or architectural changes; direct execution for clear, small fixes.
- In CI use
claude -p(non-interactive), with--output-format jsonfor machine-readable results.
Practise 24 Claude Code Configuration & Workflows questions →
Prompt Engineering & Structured Output 20%
- Explicit categorical criteria beat vague instructions such as "be conservative".
- Few-shot examples are the strongest lever for consistent format and judgement.
- tool_use with a JSON schema removes syntax errors but not semantic ones. Make fields nullable to stop fabrication.
- Validation-retry loops fix format errors, not information that is missing from the source.
- Message Batches API: about half the cost, up to 24h, no multi-turn tools. Never use it for blocking workflows.
Practise 24 Prompt Engineering & Structured Output questions →
Context Management & Reliability 15%
- Keep a persistent "case facts" block so amounts, dates and IDs survive summarisation.
- Escalate when a customer asks for a human, on policy gaps, or when stuck. Not on sentiment or self-reported confidence.
- Propagate errors with context and partial results. Never silently return empty "success".
- Validate accuracy per document type and field, not just the overall average.
- Preserve claim-to-source mappings and show both values when credible sources conflict.
Practise 18 Context Management & Reliability questions →
The six exam scenarios
Each exam draws 4 of these 6 scenarios at random and frames about 15 questions inside each one, so be ready for all six. Our mock exam works the same way.
An agent handles returns, billing disputes and account problems through tools such as customer lookups and refunds. It should resolve most cases on first contact and hand the rest to people with a clear summary.
A coordinator splits research questions across specialist subagents for searching, analysing and writing. It then combines their findings into a cited report.
Engineers use agents with built-in tools and MCP servers to explore codebases, look things up and automate routine development tasks.
A team uses Claude Code day to day to write and change code, with shared CLAUDE.md files, rules, skills and commands keeping everyone consistent.
Claude Code runs inside build pipelines to review pull requests, generate tests and report findings in a form other tools can read.
A pipeline pulls structured fields out of invoices, contracts and other documents, validating the results before they reach downstream systems.
Three golden rules
1. Trace failures to their origin. Symptoms show up downstream; the cause usually lives upstream (often in the coordinator or the prompt you passed).
2. High stakes means programmatic. If a single failure costs money, breaks compliance or risks security, enforce it with a hook or gate, not a prompt.
3. Choose the proportionate fix. Better tool descriptions before routing classifiers; community MCP servers before custom builds; scoped access before full access.
Further practice
For a second set of questions written by someone who has sat the exam, Matthew Purcell publishes an independent 60-question practice set organised by scenario, with answers and rationales. It's a useful complement to the mock exams here.
Paul Larionov's open-source Claude Certified Architect study guide walks through every domain in depth, and his practical test adds more questions (repository, with translations in many languages).
Try three sample questions
Pick your answer, then open the reveal. The full bank has 120 questions.
if response.content[0].type == 'text': return. What is the defect?- Text and tool_use blocks can arrive together, so it exits before pending tools run
- It should exit when stop_reason equals max_tokens, not by inspecting content types
- It has no iteration cap, so long investigations cannot recover once they stall
- The system prompt never tells the model to output 'DONE' when the work is complete
Reveal answer
get_customer instead of lookup_order. Both descriptions read 'Retrieves [entity] information.' What should you do first?- Put a routing classifier in front of the agent to pick the right tool
- Merge both tools into a single
lookuptool that handles any entity - Expand both descriptions: purpose, inputs, examples and when to use each
- Add ten few-shot examples of correct routing to the system prompt
Reveal answer
- Tom is running a different model from Priya, which follows rules less well
- The conventions are in Priya's ~/.claude/CLAUDE.md, not the project's
- Tom's clone of the repository is out of date and missing recent commits
- The conventions need to be repeated in a CLAUDE.md in every subdirectory
Reveal answer
A four-week study plan
- Week 1, Domain 1 (27%). Agentic loops, hub-and-spoke orchestration, context passing, hooks and human-in-the-loop. Do Level 5 lesson 3 first.
- Week 2, Domains 2 & 3. Tool design, structured errors, MCP scoping, CLAUDE.md hierarchy, rules, skills, plan mode and CI.
- Week 3, Domains 4 & 5. Few-shot, tool_use schemas, validation-retry, batches, case facts, escalation, provenance.
- Week 4, mock exams. Take a timed mock every other day, drill your weakest domain in between, and aim for 800+ before booking.
Frequently asked
What is the pass mark for the Claude Certified Architect exam?
The exam is scored out of 1000 and 720 is widely reported as the pass mark. Check Anthropic's official exam page for the current figure before you book.
Which domain is worth the most?
Agentic Architecture & Orchestration at 27%, followed by Claude Code Configuration & Workflows and Prompt Engineering & Structured Output at 20% each, Tool Design & MCP Integration at 18%, and Context Management & Reliability at 15%.
What kind of questions are on the exam?
Scenario-based multiple choice. Each question has one correct answer and three plausible distractors, usually set inside a realistic production scenario such as a customer support agent or a CI/CD pipeline.
Are these the real exam questions?
No. Every question here is original and written to practise the same skills and decision rules. Using leaked exam content would break the exam terms and would not help you learn.
Independent practice material, not affiliated with Anthropic. Exam details such as domain weights and pass mark reflect publicly shared information at the time of writing and may change. Always confirm on Anthropic's official certification pages.