720 is not 72%. That single mistake is repeated across most study material for the Claude Certified Architect – Foundations exam. Here is a free 20-question mock written from the guide that says otherwise.

The mock: 20 questions

Weighted like the real blueprint. Two items are multiple-response, as on the exam. No signup, nothing stored, nothing sent anywhere. Check each answer as you go, then score yourself at the end.

1. 1Tool Design & MCP Integration

   An agent has grown from five to eighteen tools and selection reliability has fallen. Which two remedies align with the guide? (Select two.)

   • Restrict the agent's tool set to those relevant to its role
   • Add few-shot examples covering all eighteen tools
   • Increase the model tier so it can reason over more options
   • Replace overly generic tools with constrained alternatives that validate their inputs
   Check answer

2. 2Claude Code Configuration & Workflows

   Which two locations hold custom slash commands, and how do they differ? (Select two.)

   • .claude/rules/ — path-scoped and conditionally loaded
   • CLAUDE.md — project-scoped and always loaded
   • ~/.claude/commands/ — user-scoped and personal
   • .claude/commands/ — project-scoped and shared via version control
   Check answer

3. 3Agentic Architecture & Orchestration

   Which task is best served by dynamic adaptive decomposition rather than a fixed pipeline?

   • Running the same four-section audit on every microservice in a repo
   • Generating API documentation from a stable OpenAPI specification
   • Converting one hundred files from JavaScript to TypeScript
   • Investigating why an intermittent production failure occurs, where each finding changes what to check next
   Check answer

4. 4Agentic Architecture & Orchestration

   Which best describes fork-based session management in the context of subagent work?

   • Isolating a skill's output from the main conversation
   • Exploring divergent approaches from a shared analysis baseline
   • Distributing one agent's workload across several processes
   • Persisting agent state to disk for crash recovery
   Check answer

5. 5Agentic Architecture & Orchestration

   You need to both normalise inconsistent tool output formats AND block a policy-violating action. Which combination is correct?

   • A single PostToolUse hook handling both concerns
   • A tool call interception hook for normalisation, and a PostToolUse hook for blocking
   • A PostToolUse hook for normalisation, and a tool call interception hook for blocking
   • A single tool call interception hook handling both concerns
   Check answer

6. 6Agentic Architecture & Orchestration

   When escalating mid-conversation to a human agent who cannot see the transcript, what should the handoff include?

   • The full untruncated conversation transcript so nothing is lost
   • Customer ID, root cause analysis, refund amount, and recommended action
   • A link to the agent's session so the human can resume it directly
   • A confidence score explaining why the agent could not resolve the case
   Check answer

7. 7Agentic Architecture & Orchestration

   Which scenario is the weakest justification for a hook?

   • Ensuring agent replies consistently use British rather than American spelling
   • Blocking any refund above the policy threshold
   • Guaranteeing identity verification precedes financial operations
   • Normalising inconsistent data formats returned by three different backends
   Check answer

8. 8Agentic Architecture & Orchestration

   Which pairing of task to decomposition strategy is correct in both cases?

   • Both → prompt chaining, since predictability aids CI
   • Fixed four-aspect PR review → prompt chaining; open-ended production incident investigation → dynamic adaptive decomposition
   • Fixed four-aspect PR review → dynamic adaptive decomposition; open-ended incident investigation → prompt chaining
   • Both → dynamic adaptive decomposition, since real inputs always vary
   Check answer

9. 9Tool Design & MCP Integration

   Which distinction does the guide say prevents wasted retry attempts?

   • Returning a stack trace so the agent can diagnose the cause
   • Returning structured metadata that separates retryable from non-retryable errors
   • Returning HTTP status codes rather than prose messages
   • Returning errors asynchronously so the agent can continue
   Check answer

10. 10Tool Design & MCP Integration

    A refund is refused because it violates a policy limit. Which error shape lets the agent respond appropriately without wasting retries?

    • isError true, errorCategory 'transient', isRetryable true, with a retry_after hint
    • A plain exception message containing the policy text for the model to parse
    • An empty successful result so the agent moves on to alternative resolutions
    • isError true, errorCategory 'business', isRetryable false, plus a customer-friendly explanation
    Check answer

11. 11Tool Design & MCP Integration

    An Edit call fails because the anchor text appears in six places. What is the documented fallback?

    • Use Glob to locate a file where the anchor is unique
    • Call Edit repeatedly until all six are updated
    • Delete the file and Write a fresh version from memory
    • Read the file, then Write it back with the change applied
    Check answer

12. 12Claude Code Configuration & Workflows

    A monorepo has packages with genuinely different standards, and each package maintainer knows their own domain. Which mechanism keeps CLAUDE.md modular while letting maintainers own their rules?

    • @import syntax referencing the specific standards files relevant to each package
    • A user-level CLAUDE.md per maintainer describing their own package
    • One root CLAUDE.md listing every package's standards under headings
    • Duplicating the full standards file into every package directory
    Check answer

13. 13Claude Code Configuration & Workflows

    Your automated review re-reports the same issues on every push, even after they are fixed. What is the recommended fix?

    • Cache the previous run's output and diff it before posting comments
    • Reduce review frequency to once per pull request rather than per push
    • Only review files changed since the last successful run
    • Give the run the findings from the review before it, and tell it to report only what is new or still unaddressed
    Check answer

14. 14Claude Code Configuration & Workflows

    Your migration script mishandles null values. Which correction does the guide describe as most effective?

    • Describe the null-handling policy in more precise prose
    • Ask the model to add defensive null checks throughout
    • Add the null case to CLAUDE.md as a standing rule
    • Provide a specific test case with example input containing nulls and the expected output
    Check answer

15. 15Prompt Engineering & Structured Output

    Before submitting 50,000 documents to the batch API, what does the guide recommend?

    • Split the submission into batches of exactly 100 for reliability
    • Pre-validate every document against the schema before submission
    • Run the first 5,000 synchronously to warm the model's behaviour
    • Refine the prompt on a sample set to maximise first-pass success and reduce resubmission costs
    Check answer

16. 16Prompt Engineering & Structured Output

    Why does the guide say high false positive rates in one category are especially damaging?

    • They cause the model to overfit to that category over time
    • They undermine developer confidence in the accurate categories too
    • They increase API costs disproportionately
    • They delay the pipeline beyond acceptable build times
    Check answer

17. 17Prompt Engineering & Structured Output

    Your per-file review passes are working well — findings are consistent and specific — but the pipeline keeps missing bugs where one file's change breaks an assumption another file relies on. How should you extend the architecture?

    • Ask the session that produced the per-file findings to re-read them and infer the cross-file effects
    • Re-check only the files whose per-file pass returned the highest-severity findings
    • Add a separate integration pass over the full change set, run after the per-file passes and focused on cross-file data flow
    • Have each per-file pass also load every file that file imports, so cross-file context is present throughout
    Check answer

18. 18Prompt Engineering & Structured Output

    An engineer believes that because their schema is strict, extracted invoices cannot contain incorrect totals. What is the error in this reasoning?

    • Strict schemas are only enforced when tool_choice is forced
    • Strict schemas eliminate syntax errors but not semantic errors, such as line items that do not sum to the stated total
    • Strict mode applies to input parameters, not to extracted values
    • Schema validation runs after the response is parsed, so it cannot reject values
    Check answer

19. 19Context Management & Reliability

    Each subagent in a research pipeline returns its findings with source URLs attached. The pipeline summarises each subagent's output before synthesis, and the final report states claims without saying which source supports them. Where was the attribution lost?

    • At the summarisation step, which compressed the findings without carrying the claim-to-source mapping through
    • In the report template, which has no field for citations
    • At synthesis, which merged findings from subagents using different citation formats
    • At collection, because the subagents were never asked to record their sources
    Check answer

20. 20Context Management & Reliability

    During analysis a document is found to contain two conflicting values for the same figure. What should the document analysis agent return?

    • A null value with an error flag, since the document is unreliable
    • The value appearing later in the document, as it likely supersedes the earlier
    • A single reconciled value with the reasoning recorded internally
    • Both values, explicitly annotated as conflicting, letting the coordinator decide how to reconcile
    Check answer

Score my mock

Found a question that felt unfair or wrong? Tell me which one — it gets fixed.

Prefer paper? The same 20 questions and the 8 misconceptions as free PDFs (email required).

If this helped, star the repo — it is how other people find it.