Foundations Mock Exam 1 — 60 Scenario Questions, Timed (Q529–Q588)

A full-length CCAR-F practice simulation, built like the official exam: 60 questions in 120 minutes, under 4 scenarios drawn from the bank of 6, 15 questions per scenario. This mock uses Customer Support Resolution Agent · Code Generation with Claude Code · Multi-Agent Research System · Structured Data Extraction. Every question was written for this guide against a task statement of the official Claude Certified Architect – Foundations Exam Guide (v1.0, July 2026), and the domain counts follow its blueprint. 4 questions are multiple-response and say “Select TWO”; the rest say “Select ONE”. This is not an official Anthropic or Pearson VUE exam. It does not reproduce, reconstruct or recall any live exam item, and the guide’s sample questions are used only as a model of style. Other full mocks: Mock 2 · Mock 3 · Mock 4.

How to read your result. The runner reports a practice score: raw questions correct out of 60, a per-domain breakdown, the elapsed time and the unanswered count. Unanswered questions count as incorrect. The live exam reports a scaled score (pass mark 720 on 100–1,000); this page does not convert a raw score into a scaled score, because no public conversion exists. A useful study heuristic: aim for 80% here, with no domain under 65%, on a mock you have not seen before. Untimed practice by domain lives in Practice Questions and Concept Drills.

Foundations mock exam 1 · 60 questions (Q529–Q588)

Scenario 1 · Customer Support Resolution Agent

D5 · Context & Reliability

Q529. You are drafting the escalation section of Larkspur's system prompt. Reviewers found the agent handing calm customers with routine returns to humans while it kept handling cases it should have passed on. Which two situations should the section list as reasons to escalate? (Select TWO.) A) The customer explicitly asks to talk to a person, even when the agent believes it could resolve the issue itself in its next message. B) The agent's own self-reported confidence in its planned resolution drops below 70%, suggesting the case is probably beyond its abilities. C) The customer requests something the refund policy does not cover, such as reimbursing a hotel night spent waiting for a replacement tent. D) The sentiment score of the customer's latest message falls below a set threshold, showing a level of frustration a person may handle better.

D4 · Tools & MCP

Q530. Policy answers come from 60 documents behind a search_policies tool. Logs show the agent averages five exploratory search_policies calls per chat just to learn which documents exist, adding about eight seconds per reply. The documents change monthly. Which change best reduces these exploratory calls? (Select ONE.) A) Cap search_policies at two calls per conversation in the harness, so the agent stops exploring and answers from whatever policy text it has found by then. B) Expose the policy catalog, with titles, topics and short summaries, as an MCP resource the agent sees up front, then fetch only the documents it needs. C) Paste all 60 policy documents into the system prompt so the agent never needs to search, and update the prompt each time a policy changes. D) Replace search_policies with a fine-grained keyword index tool, so that each search returns more precise matches on the first attempt.

D3 · Prompting & Structured Output

Q531. After each chat, the agent must write a wrap-up record to the CRM with category, order IDs, resolution and refund amount. You currently ask for JSON in the final message, and 3% of records fail to parse because of prose before the JSON or a missing bracket. Which approach fixes this most reliably? (Select ONE.) A) Add a stronger instruction such as 'respond with JSON only, no other text', and include one complete example record in the system prompt. B) Keep the free-text JSON, but add a regular-expression parser that strips prose and repairs missing brackets before the record is saved. C) Lower the temperature to zero for the wrap-up turn, so that the model produces the JSON format more consistently across all chats. D) Define a record_resolution tool whose input_schema is the CRM record, force it with tool_choice at wrap-up, and read the fields from the tool_use input.

D4 · Tools & MCP

Q532. Larkspur added a check_store_credit tool with a detailed description covering balances, expiry dates and gift-card credit. Yet for 'how much store credit is on my account?' the agent still calls get_customer and answers from its sparse loyalty field. The system prompt contains: 'For any question about the customer's account, call get_customer first.' What should you change first? (Select ONE.) A) Remove the loyalty field from get_customer's output, so the agent must look for another tool whenever a customer asks about stored credit. B) Make check_store_credit's description much longer, with more example questions, so that it outweighs the get_customer rule in the system prompt. C) Add a routing classifier before the agent that detects store-credit questions and calls check_store_credit directly, bypassing the model's choice. D) Rewrite the system prompt rule to cover identity and profile questions only, and point store-credit and balance questions to check_store_credit.

D1 · Agentic Architecture

Q533. Larkspur's agent loop ends the turn whenever the model's response contains a text block. Logs show many replies such as 'Let me pull up that order for you' arriving together with a lookup_order tool_use block; the loop returns that sentence to the customer and never runs the tool. Which change to the loop control fixes this? (Select ONE.) A) Continue the loop while stop_reason is "tool_use", run each requested tool and append its result, and finish only when stop_reason is "end_turn". B) Keep the text check but require at least three loop iterations before returning, so short acknowledgement sentences such as 'one moment' no longer end the turn early. C) Search the text block for completion phrases such as 'anything else?' or 'resolved', and end the turn only when one of those phrases appears in the reply. D) Instruct the model in the system prompt never to write text in the same response as a tool call, so a text block reliably signals that the turn is complete.

def run_turn(client, messages, tools, execute_tool):
    while True:
        response = client.messages.create(model=MODEL, max_tokens=1024, tools=tools, messages=messages)
        messages.append({"role": "assistant", "content": response.content})
        if response.stop_reason != "tool_use":
            return response  # end_turn: the model has finished
        results = []
        for block in response.content:
            if block.type == "tool_use":
                output = execute_tool(block.name, block.input)
                results.append({"type": "tool_result", "tool_use_id": block.id, "content": output})
        messages.append({"role": "user", "content": results})

D5 · Context & Reliability

Q534. A customer gives only the name 'Jordan Ellis'. get_customer returns three active accounts with that name in different states. The agent currently picks the account with the most recent order and continues, and one such guess led to a refund on a stranger's order. What should the agent do? (Select ONE.) A) Escalate to a human every time get_customer returns more than one match, since identity questions are outside the agent's scope. B) Ask the customer for another identifier, such as the order number or the account email, and continue only when exactly one account matches. C) Pick the most recent account but block process_refund on it until the customer confirms the order total in a later message. D) Choose the account whose recent orders best match the products the customer mentioned, and confirm that choice with the customer in the closing summary.

D1 · Agentic Architecture

Q535. The system prompt tells Larkspur's agent never to refund more than $300 per order without a human. An audit of 40,000 chats finds 52 process_refund calls above that limit, each after a customer pushed back several times. Finance requires zero such refunds. What should you implement? (Select ONE.) A) A PostToolUse hook that reads each completed process_refund result and lists every refund above $300 in a daily finance report so the team can claw it back later. B) Repeat the $300 limit at the end of the system prompt in capital letters, and add few-shot examples of the agent politely declining refunds after repeated pushback. C) A hook that intercepts every outgoing process_refund call, blocks any amount above $300, and redirects the case to escalate_to_human with the amount attached. D) A classifier that scores pushback in each customer message and switches the agent to a stricter system prompt once the score is high.

REFUND_LIMIT = 300

async def refund_limit_hook(input_data, tool_use_id, context):
    if input_data["tool_name"] != "mcp__support__process_refund":
        return {}
    amount = input_data["tool_input"].get("amount", 0)
    if amount > REFUND_LIMIT:
        return {"hookSpecificOutput": {
            "hookEventName": "PreToolUse",
            "permissionDecision": "deny",
            "permissionDecisionReason": f"Refund {amount} exceeds {REFUND_LIMIT}. Call escalate_to_human with this amount."}}
    return {}

D1 · Agentic Architecture

Q536. A customer writes that a camp stove arrived dented, that they were charged twice for it, and that their account still shows an old email address. The agent currently handles only the first problem it notices and ends the chat, and 31% of multi-issue customers contact support again. Which design should you adopt? (Select ONE.) A) Have the agent split the message into separate issues, investigate each using the shared customer context, then give one combined resolution covering all three. B) Ask the model to rank the three issues by urgency, resolve only the most urgent one fully, and mention the others in its closing message. C) Escalate every message that mentions more than one problem to a human, since multi-issue requests are too complex for autonomous handling. D) Resolve the dent now, then invite the customer to open a new chat for the double charge and a third chat for the email change.

D4 · Tools & MCP

Q537. During a database failover, lookup_order returned an empty orders list instead of an error for 20 minutes. The agent told 140 customers that no orders existed on their accounts, and several placed duplicate orders for items they had already bought. What change to lookup_order prevents this? (Select ONE.) A) Escalate to a human every conversation in which lookup_order returns no orders, since an empty result might hide a backend outage. B) Add a system prompt rule telling the agent to double-check with get_customer whenever lookup_order returns an empty orders list, before it tells the customer anything. C) Return an error marked transient and retryable when the order database is unreachable, and keep an empty list only for successful queries that find nothing. D) Cache each customer's last known orders inside the tool, and return the cached list whenever the database is unreachable, without informing the agent of the outage.

D4 · Tools & MCP

Q538. process_refund returns {"error": "Operation failed"} both when the payment gateway times out and when an order is outside the 30-day window. Logs show the agent retrying window violations three times, then telling customers 'a temporary issue occurred, please try later.' How should the tool report failures? (Select ONE.) A) Tell the agent to stop retrying after the first failure of any kind and to escalate to a human every time any refund call returns an error of any type. B) Keep one generic message but add a numeric status code to it, and list in the system prompt which codes the agent may retry and which it may not. C) Return isError true with an errorCategory such as transient or business, an isRetryable flag, and a customer-safe explanation of the failure. D) Retry gateway timeouts inside the tool up to five times with backoff, and return the same 'Operation failed' text once every attempt has been used up.

{
  "timeout_example": {"isError": true, "content": [{"type": "text", "text": "{\"errorCategory\": \"transient\", \"isRetryable\": true, \"description\": \"Payment gateway timed out after 10 s.\"}"}]},
  "policy_example": {"isError": true, "content": [{"type": "text", "text": "{\"errorCategory\": \"business\", \"isRetryable\": false, \"description\": \"Order 5521 was delivered 41 days ago; the return window is 30 days.\"}"}]}
}

D1 · Agentic Architecture

Q539. lookup_order reads from two warehouse systems. The legacy one returns delivered_at as a Unix timestamp; the newer one returns an ISO 8601 string with a time-zone offset. The agent misjudges Larkspur's 30-day return window on about 9% of legacy orders and on none of the others. What is the most reliable fix? (Select ONE.) A) Add a PostToolUse hook that converts every delivered_at value to one ISO 8601 UTC format before the result reaches the model, whatever its source. B) Ask the model to state the delivery date it computed before deciding eligibility, so reviewers can catch conversion mistakes in the logs. C) Move the agent to a larger model with stronger arithmetic, since the errors come from converting epoch seconds into dates during reasoning. D) Add system prompt instructions that explain how to convert Unix timestamps into calendar dates, with two worked examples of the 30-day calculation.

D1 · Agentic Architecture

Q540. Human agents who receive Larkspur escalations spend about six minutes per case re-asking customers for order numbers and what was already tried. The escalate_to_human call currently sends one free-text field, for example 'Customer needs help with a refund.' What change would most reduce this rework? (Select ONE.) A) Attach the full chat transcript and every raw tool result as files on the escalation ticket, so human agents can read the complete history themselves before replying. B) Tell the agent in the system prompt to write a more detailed escalation message, and review a sample of escalation texts each week for quality. C) Route all escalations to a senior queue, where experienced agents can resolve cases quickly without needing background from the conversation. D) Make escalate_to_human require structured fields: customer ID, order IDs, root cause found, actions already taken, amount at stake and a recommended next step.

{
  "name": "escalate_to_human",
  "description": "Hand the case to a human agent who cannot see this chat. Fill every field from facts confirmed in the conversation.",
  "input_schema": {
    "type": "object",
    "properties": {
      "customer_id": {"type": "string"},
      "order_ids": {"type": "array", "items": {"type": "string"}},
      "root_cause": {"type": "string"},
      "actions_taken": {"type": "array", "items": {"type": "string"}},
      "amount_at_stake": {"type": ["number", "null"]},
      "recommended_action": {"type": "string"}
    },
    "required": ["customer_id", "order_ids", "root_cause", "actions_taken", "recommended_action"]
  }
}

D5 · Context & Reliability

Q541. In long chats, Larkspur summarizes older turns to save context. A review finds that after summarization the agent sometimes quotes the wrong refund amount: a customer promised $84.20 on order 5521 was later offered 'a partial refund of about $80.' Which design keeps such details reliable? (Select ONE.) A) Summarize more often with shorter summaries, so that each summary covers fewer turns and has less room to drop numerical details. B) Stop summarizing and send the complete transcript every time, accepting slower and costlier requests so that no detail can ever be lost. C) Instruct the summarizer to be more careful with numbers, and add a final step asking the model to confirm any amount it plans to quote. D) Keep a persistent case-facts block of order IDs, amounts, dates and promises made, included in every request outside the summarized history.

D1 · Agentic Architecture

Q542. Yesterday a customer began a damaged-tent claim; the chat ended while the agent waited for photos. Overnight the warehouse shipped a replacement, so lookup_order now shows a new status. Today the customer returns with photos. Your application can resume yesterday's session, whose history still holds the old lookup_order results. What should it do? (Select ONE.) A) Resume yesterday's session and move to a model with a larger context window, so the old order data and today's photos both fit without summarizing. B) Start a new session seeded with a structured summary of yesterday's claim, noting the order changed overnight, so the agent calls lookup_order again. C) Fork yesterday's session into two branches, one assuming the replacement shipped and one assuming it did not, then keep the branch the customer confirms. D) Resume yesterday's session unchanged, so the agent has the full conversation, including the earlier lookup_order results, before it reviews the photos.

D4 · Tools & MCP

Q543. To reuse company tooling, the platform team connected Larkspur's agent to 19 MCP tools, including inventory forecasting, marketing coupons and supplier messaging. Since then, wrong-tool calls rose from 2% to 11%, mostly issue_coupon calls where process_refund was required. What should you do? (Select ONE.) A) Set tool_choice to "any" so the agent must always call one of the 19 tools, which stops conversational guesses about its next action. B) Keep all 19 tools, and add a numbered priority list to the system prompt that tells the agent which tools to try first for refund cases. C) Give the support agent only the tools its role needs, such as get_customer, lookup_order, process_refund and escalate_to_human, and no others. D) Move to a larger model with more reasoning capacity, since choosing correctly among 19 tools is a load the current model cannot carry.

Scenario 2 · Code Generation with Claude Code

D2 · Claude Code

Q544. Quarry's root CLAUDE.md now includes a 200-line release-notes procedure that engineers use about twice a month. Every session loads it, and engineers report Claude sometimes starts drafting release notes when they only asked for a commit message. Where should this procedure live? (Select ONE.) A) In a project skill at .claude/skills/release-notes/SKILL.md, committed to git, so it loads when the task calls for it and stays shared. B) In a .claude/rules/ file with a paths pattern matching CHANGELOG.md, so that the procedure loads only when someone opens or edits the changelog file. C) In the root CLAUDE.md still, but placed under a heading that tells Claude to use the procedure only when release notes are explicitly requested. D) In each engineer's ~/.claude/CLAUDE.md, so it no longer loads for the whole team and each person can keep whichever version of the procedure they prefer.

D2 · Claude Code

Q545. An engineer who has never built a cache asks Claude Code to add one in front of Quarry's slow route-pricing service. The engineer is unsure what should happen when carrier rates change, how stale a price may be, or what to do when the cache store is down. What is the best way to start? (Select ONE.) A) Use plan mode only to list the files that call the pricing service, then let Claude add the cache directly in each of those files in one pass. B) Ask Claude to implement a standard read-through cache immediately, then review the resulting diff to see which invalidation choices it made on its own. C) Write a CLAUDE.md rule that every cache must use a 60-second expiry, so Claude applies a safe default to the pricing cache without further discussion. D) Ask Claude to interview the engineer first about invalidation triggers, staleness limits and failure modes, then implement the design they agree on.

D4 · Tools & MCP

Q546. An engineer new to Quarry asks Claude Code how a shipment-delay alert reaches the mobile app. Claude's first plan is to Read all 1,900 files in the monorepo before answering, which would take a long time and fill the context. Which exploration approach should Claude use instead? (Select ONE.) A) Use Bash to concatenate all source files into one large file, then Read that single file so the whole codebase is loaded at once. B) Use Glob with the pattern **/*.ts to list every TypeScript file, then Read them in alphabetical order until the alert path becomes clear. C) Use Glob to find files whose names contain 'alert', and answer from those files alone without following any imports to other modules. D) Use Grep to find where the delay-alert event name or message appears, then Read those entry files and follow their imports step by step.

D2 · Claude Code

Q547. Quarry's root CLAUDE.md has grown to 1,100 lines covering React, API, Terraform and data-pipeline standards. Engineers in the infra package say sessions often apply React advice to Terraform work. The maintainers of each package know which standards apply to their area. How should you restructure the instructions? (Select ONE.) A) Split the standards into separate files, and give each package a CLAUDE.md that uses @import to include only the files its maintainers choose. B) Convert every standard into a skill that engineers invoke by hand, so that no conventions load until someone decides they are needed. C) Move the whole file into each engineer's ~/.claude/CLAUDE.md, so everyone can delete the sections that do not concern the package they work on. D) Keep one root file but reorder it so the most frequently used standards appear first, where Claude pays the most attention during a long session.

D2 · Claude Code

Q548. An engineer asks Claude Code to write a function that turns carrier route codes into display labels. After three prose explanations, the results still vary: sometimes 'NYC-BOS-EXP' becomes 'NYC to BOS (Express)', other times 'Express: NYC–BOS'. What is the most effective next step? (Select ONE.) A) Rewrite the prose description a fourth time with more precise wording, and ask Claude to restate the rules before writing the function. B) Give two or three concrete input and expected output pairs, including an edge case such as a code without a service suffix, and ask Claude to match them. C) Ask Claude to generate several alternative implementations, then pick the one whose output style the team likes best for the display. D) Switch to plan mode so Claude can explore the codebase for existing label formats before it writes a new version of the function.

D1 · Agentic Architecture

Q549. Yesterday an engineer ran a named Claude Code session, billing-refactor, that analyzed Quarry's twelve-file invoicing module. Overnight a teammate merged changes to two of those files, invoice_tax.ts and rounding.ts. Most of yesterday's analysis is still valid. How should the engineer continue? (Select ONE.) A) Resume with claude --resume billing-refactor and keep working from yesterday's tool results for all twelve files, since most analysis is still valid. B) Start a brand-new session and ask Claude to analyze all twelve files from scratch, since any merged change makes the earlier session unreliable. C) Resume with claude --resume billing-refactor and tell Claude that invoice_tax.ts and rounding.ts changed, so it re-reads and re-analyzes those two. D) Fork yesterday's session with --fork-session, so the original analysis stays safe while the new branch continues without mentioning the changes.

claude --resume billing-refactor
> Since yesterday, invoice_tax.ts and rounding.ts were changed by a merged PR.
> Re-read those two files, update your analysis of tax and rounding, then continue the refactor plan.

D2 · Claude Code

Q550. Quarry's pull-request review job already runs Claude Code with -p. Two complaints remain: each new commit reposts findings raised on earlier commits, and the comment-posting script fails on about 5% of runs because findings arrive as loosely formatted text. Which two changes address these problems? (Select TWO.) A) Run with --output-format json and --json-schema describing file, line, severity and message, so the script reads validated structured findings. B) Resume the previous commit's review session with --resume, so Claude remembers its earlier findings and naturally avoids repeating them. C) Add a CLAUDE.md rule asking Claude to format every finding as a JSON object, so the posting script receives parseable output on each run. D) Pass the findings already posted on the pull request into the review prompt, and tell Claude to report only issues that are new or still unaddressed.

gh pr view "$PR" --comments --json comments > prior.json
gh pr diff "$PR" | claude -p \
  --output-format json \
  --json-schema "$(cat .github/review-schema.json)" \
  "Review this diff. Prior findings are in prior.json. Report only issues that are new or still unaddressed."

D2 · Claude Code

Q551. Before a refactor, an engineer asks Claude Code to find every place the API reads feature flags. Discovery reads over 100 files, and the main session's context fills with search output; by the implementation phase Claude has lost the constraints agreed at the start. Which approach best protects the main session during discovery? (Select ONE.) A) Run discovery in direct execution but ask Claude to print only file names, then reread every file in full during the implementation phase. B) Raise the session's maximum output tokens, so Claude can write longer answers that keep both the search output and the agreed constraints. C) Restart the session after discovery and paste the whole search output into the first message, so the new context begins complete. D) Delegate discovery to the Explore subagent, which reads files in its own context and returns a concise summary of where flags are read.

D2 · Claude Code

Q552. Quarry created .claude/rules/api-handlers.md with paths: ["services/api/.ts"]. Engineers say the error-handling conventions apply when they edit services/api/index.ts but not when they edit services/api/orders/create.ts, and /memory during that second edit does not list the rule. What is the most likely fix? (Select ONE.) A) Rename the file to api-handlers.yaml, because rule files with paths frontmatter must use a YAML file extension to be parsed. B) Copy the rule's content into the root CLAUDE.md, because path-scoped rules are loaded only once at session start and never during later edits. C) Change the pattern to "services/api/**/.ts", because a single asterisk matches files in that one folder and not files in nested subfolders. D) Move the rule into services/api/orders/CLAUDE.md, because path rules only take effect in the folder that directly contains the rules file itself.

D2 · Claude Code

Q553. Quarry's Terraform conventions (module naming, required tags, no inline policies) sit in the root CLAUDE.md, so every React and API session loads them too. Terraform files live under infra/, but also in services//deploy/ folders owned by each service team. How should you load these conventions only when they matter? (Select ONE.) A) Create .claude/rules/terraform.md with paths frontmatter such as "**/.tf", so the conventions load when Claude works with Terraform files anywhere. B) Create a /terraform slash command containing the conventions, and ask engineers to run it before they begin any infrastructure change. C) Keep the conventions in the root CLAUDE.md, but mark the section 'Terraform only' so Claude ignores it during React and API sessions. D) Create infra/CLAUDE.md with the conventions, and copy that file into every services/*/deploy/ folder so each Terraform location has its own local copy.

# .claude/rules/terraform.md (the frontmatter block at the top of the file)
paths:
  - "**/*.tf"
# Markdown body below the frontmatter, for example:
#   - Module names: <team>-<purpose>, lowercase, hyphenated
#   - Every resource sets tags: owner, cost_center, env
#   - No inline IAM policies; reference policy modules

D2 · Claude Code

Q554. Quarry's nightly CI job asks Claude Code to generate tests for recently changed modules. Engineers discard about 40% of the output: many tests repeat scenarios already covered, and others use hand-built mocks although the repo has shared fixtures in test/fixtures. Which change most improves the generated tests? (Select ONE.) A) Move the job to the Message Batches API so it can process every module overnight with more time to think about each test. B) Generate twice as many tests per module, and let a second pass delete any test whose name matches an existing test in the repository. C) Include each module's existing test files in context, and document testing standards, valuable test criteria and the shared fixtures in CLAUDE.md. D) Ask in the job's prompt for 'high-quality, non-duplicate tests only', and lower the number of tests requested per module from ten to five.

D5 · Context & Reliability

Q555. During a two-day investigation of a memory leak across Quarry's API services, Claude Code's answers start to drift: it now mentions 'a typical connection-pool pattern' instead of the PoolManager class it identified on day one, and it contradicts earlier findings. What should the engineer do to keep findings reliable for the rest of the investigation? (Select ONE.) A) Keep the session open the whole time, so all of day one's tool results remain in context and Claude can look back at them when needed. B) Have Claude record confirmed findings, class names and file paths in a scratchpad file, and refer to that file for each later question. C) Switch to a model with a larger context window, so the two days of exploration fit without the older details becoming less reliable over time. D) Ask Claude to reread every file it examined on day one at the start of each morning, so the original details are fresh in the session again.

D2 · Claude Code

Q556. An engineer must replace Quarry's in-house date utilities with a standard date library. The utilities are used in about 60 files across the front end and API, two areas depend on undocumented time-zone behavior, and there are at least two reasonable migration orders. How should the engineer use Claude Code? (Select ONE.) A) Use direct execution and ask Claude to migrate all 60 files in one pass, then rely on the existing test suite to catch any time-zone regressions afterwards. B) Ask Claude to write a detailed migration plan into CLAUDE.md first, so that every later session follows the plan automatically without needing plan mode. C) Start in plan mode to explore usages and the time-zone dependencies, agree on a migration plan, then use direct execution to apply it in stages. D) Use direct execution one file at a time, letting Claude decide each file's changes as it goes, so the engineer can review every small diff before moving on.

D2 · Claude Code

Q557. Two Quarry engineers run the same /fix-lint command on the same branch and get different results: one gets the team's Result-type error handling, the other does not. Nothing in the command file mentions error handling, and you suspect a configuration difference between their machines. What is the most effective first step to find the cause? (Select ONE.) A) Ask both engineers to start their sessions with --bare, so neither loads personal files and both receive the same project configuration. B) Run /memory in both engineers' sessions to see which CLAUDE.md and memory files each one has loaded, then compare where the rule comes from. C) Copy the Result-type rule into the /fix-lint command file, so the command always carries the convention whichever memory files each session happens to load. D) Reinstall Claude Code on the second engineer's machine to clear any corrupted local cache that might be hiding the project instructions.

D2 · Claude Code

Q558. Quarry's /dependency-audit skill reads every package.json and lockfile, then prints a long analysis before a short final summary. Engineers run it in the middle of other tasks, and afterwards their sessions lose track of earlier work because the audit output fills the context. They need only the summary. What change best fixes this? (Select ONE.) A) Tell engineers to run /compact right after each audit, so the verbose analysis is summarized before they continue with their earlier task. B) Move the audit steps into the root CLAUDE.md, so they are always available and Claude can run the audit without invoking any skill. C) Add context: fork to the skill's SKILL.md frontmatter, so it runs in an isolated subagent context and only its final result returns. D) Add an instruction in the skill body asking Claude to keep the analysis short and to print only the most important packages it finds.

---
name: dependency-audit
description: Audit package.json and lockfiles for outdated or risky dependencies. Returns a short summary.
context: fork
agent: general-purpose
---
Read every package.json and lockfile in the repo. Compare versions with the policy in docs/deps.md.
Return only: a table of risky packages, why, and the recommended action.

Scenario 3 · Multi-Agent Research System

D1 · Agentic Architecture

Q559. After a three-hour analysis session on parking pricing, a Northwind analyst wants to compare two report framings, one built around revenue and one around congestion. Both must start from the same analysis, and work on one framing must not influence the other. Which approach fits? (Select ONE.) A) Start two new sessions and paste only the final report outline into each, so both framings begin from an identical clean context. B) Continue in the single session, writing the revenue framing first and then asking Claude to set it aside before writing the congestion framing. C) Resume the analysis session twice at the same time with --resume, so both conversations append to one shared session history. D) Use fork_session from the analysis session to create two independent branches, one per framing, each starting from the shared analysis.

D1 · Agentic Architecture

Q560. Northwind's coordinator loop stops after a fixed 8 iterations. On broad briefs the coordinator is often still delegating when the cap hits, so the report agent receives partial findings; simple briefs finish in three iterations. How should the loop decide that the coordinator is done? (Select ONE.) A) Raise the cap to 20 iterations, so broad briefs have enough room to finish while simple briefs still end well before reaching the limit. B) Continue while stop_reason is "tool_use" and stop on "end_turn", keeping a much higher iteration limit only as a guard against runaway loops. C) Stop when the coordinator's response text contains the phrase 'research complete', which the system prompt instructs it to write once every city is done. D) Make the cap depend on the size of the brief, allowing two iterations per city, so larger briefs automatically receive more iterations.

D5 · Context & Reliability

Q561. Client auditors flag two problems in Northwind briefs. Some claims can no longer be traced to any source after synthesis. And one brief called bike-share trip counts 'conflicting' when the two figures were actually from 2019 and 2024. Which two changes to subagent outputs address these problems? (Select TWO.) A) Have the synthesis agent keep only the most recent figure whenever two sources disagree, and drop older values from the brief entirely. B) Require subagents to include each figure's publication or data-collection date in their structured output, so synthesis can tell trends from conflicts. C) Require subagents to return structured claim-source mappings, with source URL, document name and supporting excerpt, that synthesis must preserve and merge. D) Ask subagents to summarize their findings in short prose paragraphs, so the synthesis agent receives less text and makes fewer mistakes.

D1 · Agentic Architecture

Q562. A city asks Northwind why bus ridership fell 18% after a 2025 route redesign. Nobody knows in advance whether the cause is service gaps, fares, remote work or bad counting data. The current coordinator always runs a fixed four-step chain built for policy comparisons. How should it decompose this request? (Select ONE.) A) Ask the coordinator to pick the single most likely cause first and research only that one, to keep the investigation narrow and fast. B) Keep the fixed four-step chain, because a predictable sequence makes the results easier to compare with the city's earlier policy briefs. C) Split the request upfront into four equal subtasks, one per suspected cause, run them in parallel, and synthesize whatever each subagent returns at the end. D) Use adaptive decomposition: investigate broadly first, then create follow-up subtasks from each step's findings, such as auditing counts if data looks off.

D1 · Agentic Architecture

Q563. Every Northwind request runs the full pipeline: search, document analysis, synthesis and report. A client question such as 'What is the current base fare in Denver?' takes six minutes and returns a four-page brief, while full policy comparisons work well. How should the coordinator change? (Select ONE.) A) Instruct the report agent to shorten its output for simple questions, so clients receive one short paragraph with a source instead of four pages of text. B) Add a separate fast pipeline that uses a smaller model for short questions, selected by a keyword classifier that runs before the coordinator. C) Keep the full pipeline but run all four subagents in parallel, so simple questions finish faster even though every agent still runs on every request. D) Have the coordinator assess each request's requirements and invoke only the subagents it needs, for example just search for a single current fact.

D5 · Context & Reliability

Q564. During a 20-city brief, three city websites return server errors all day, so the search agent cannot collect their fare data. Currently any subagent failure makes the coordinator abort the whole brief. Clients prefer an on-time brief with clearly stated gaps. How should these failures be handled? (Select ONE.) A) Have the search agent return empty results for the three cities as successful searches, so the coordinator can finish the brief on schedule. B) Have the search agent report each unreachable site as a structured failure with partial results, continue the brief, and mark coverage gaps for those cities. C) Have the search agent report 'search unavailable' for the three cities, and let the coordinator quietly drop them from the brief's scope. D) Keep aborting on any failure, but schedule an automatic retry of the whole brief later, so the three sites have another chance to respond.

D4 · Tools & MCP

Q565. Northwind wants its search agent to query a widely used academic citation index. One engineer proposes building a custom MCP server for it, although a maintained community MCP server for that index already exists. The team's own fare-history database also needs an MCP server. How should you divide the work? (Select ONE.) A) Use the maintained community server for the citation index, and build a custom MCP server only for the team-specific fare-history database. B) Use the community server for both by adding the fare-history tables to it, so the team only ever has to maintain one integration. C) Skip MCP for the citation index and have the search agent call its REST API through a generic HTTP tool, with the endpoint in the prompt. D) Build custom servers for both systems, so the team controls every tool description and can tune both integrations closely for Northwind's research work.

D5 · Context & Reliability

Q566. Traces show the document analysis agent sends every error straight to the coordinator, including brief rate-limit responses that succeed on a second attempt. The coordinator spends many turns re-delegating these tasks, and its context fills with error messages. What change fits best? (Select ONE.) A) Make the analysis agent stop the whole brief on any rate-limit response, so the operations team can investigate the document provider before work continues. B) Let the analysis agent retry transient errors with backoff, and send the coordinator only failures it cannot resolve, with attempts and partial results. C) Have the coordinator ignore every error from the analysis agent and continue the brief with whatever documents were analyzed successfully. D) Give the coordinator a larger context window, so that error messages from repeated re-delegation no longer crowd out its research findings and plans.

D4 · Tools & MCP

Q567. The document analysis agent uses a generic fetch_url tool. Traces show it sometimes follows tracking links, loads paywall login pages, and once pulled a 300 MB dataset that exhausted its context. It should only load documents the search agent returned. What change fits best? (Select ONE.) A) Keep fetch_url but add a system prompt rule listing which domains and which file sizes the agent should avoid when it fetches documents. B) Wrap fetch_url in a retry policy with a 30-second timeout and a size warning, so oversized or slow downloads fail fast and the agent moves on. C) Give the agent the web search tool as well, so it can look for a better copy of a document whenever fetch_url returns a paywall login page or a tracking redirect. D) Replace fetch_url with a load_document tool that accepts only document references from the search results, checks type and size, and returns parsed text.

D1 · Agentic Architecture

Q568. For a brief on e-scooter regulation, the coordinator sends the same broad query to two parallel search subagents to speed up collection. Their results overlap by about 70%, and the synthesis agent spends much of its context removing duplicate sources. What change best reduces the duplication? (Select ONE.) A) Have the coordinator partition the scope first, giving each search subagent distinct subtopics or source types, such as ordinances. B) Add a deduplication step to the synthesis agent's prompt, so that it removes repeated sources by URL and title before it starts comparing the findings. C) Run only one search subagent with a higher result limit, since two agents searching the same topic in parallel will always overlap heavily. D) Let the two search subagents message each other directly while they work, so each one can skip sources the other has already found.

D3 · Prompting & Structured Output

Q569. Before delivery, Northwind's report agent is asked in the same session to 'review your report for unsupported claims and errors'. It almost always approves its own drafts, yet client auditors later find unsupported claims in about one brief in twelve. What change would catch more of these issues? (Select ONE.) A) Send the report and its source mappings to a separate Claude instance without the drafting context, and have it check each claim. B) Ask the report agent to review its draft three times in a row in the same session, and approve the brief only when all three reviews agree. C) Enable extended thinking for the report agent's self-review, so it reasons longer and more carefully about each claim before it approves the draft. D) Add 'be critical and skeptical of your own work, as an auditor would be' to the report agent's system prompt before the self-review runs.

D1 · Agentic Architecture

Q570. The synthesis subagent's drafts contain claims with no citations and miss findings the document analysis agent produced. The coordinator's delegation prompt says only: 'Synthesize the findings gathered so far on transit fares.' What should the coordinator pass instead? (Select ONE.) A) A reference to the coordinator's session ID, so the synthesis subagent can load the coordinator's conversation history and read the findings there. B) A short summary of the main themes written by the coordinator, so the synthesis subagent's context stays small and focused on the big picture of the brief. C) The complete findings, structured so each claim carries its source URL, document name and page, plus the synthesis goal and quality criteria. D) A short summary of the main themes written by the coordinator, so the synthesis subagent's context stays small and focused on the big picture.

{
  "goal": "Compare base fares and fare-capping policies across the 20 cities for a city council audience.",
  "quality_criteria": ["every claim cites at least one source", "flag conflicting figures instead of choosing one"],
  "findings": [
    {"claim": "Denver base fare is 2.75", "source_url": "https://example.org/rtd-fares", "document": "RTD fare table", "page": 2, "date": "2025-07-01"},
    {"claim": "Portland caps daily fares at 5.00", "source_url": "https://example.org/trimet-hop", "document": "TriMet Hop FAQ", "page": 1, "date": "2025-03-15"}
  ]
}

D4 · Tools & MCP

Q571. The document analysis agent has two tools: read_page ('Reads a page and returns its text') and read_file ('Reads a file and returns its text'). read_file parses PDF tables; read_page handles web HTML. Traces show PDFs sent to read_page 35% of the time, losing the tables. What's the most effective first step? (Select ONE.) A) Add a system prompt rule telling the agent to use read_file for any source whose URL ends in .pdf, and to use read_page for every other source. B) Add few-shot examples of correct tool choices for PDFs and web pages to the system prompt, while keeping the two short tool descriptions unchanged. C) Remove read_page and send every source through read_file, since PDFs with fare tables are the more important source type for Northwind's briefs. D) Rename and rewrite both tools, such as fetch_web_page and parse_pdf_report, with descriptions of inputs, outputs and when to use each.

D4 · Tools & MCP

Q572. Northwind keeps a local archive of 9,000 downloaded city documents. The document analysis agent must find every council-minutes file, named like minutes-2024-03.pdf.txt across many city folders, and then find which of them mention 'fare capping'. Which built-in tool sequence fits? (Select ONE.) A) Use Grep for the word 'minutes' across the archive to find the files, then Read each matching file fully to check for fare capping. B) Use Glob with a pattern such as **/minutes-*.pdf.txt to find the files by name, then Grep those files for 'fare capping'. C) Use Bash to list all 9,000 file names into one text file, then Read that listing and pick out the minutes files by inspecting the names. D) Use Read on every folder's index file to list its documents, then use Glob with the pattern 'fare capping' to locate the relevant pages.

D1 · Agentic Architecture

Q573. Northwind's coordinator needs fare-policy findings for five cities that can be researched independently. Traces show it emits one Task call per response, waits for that result, then emits the next, so collection takes about 25 minutes. How should delegation work? (Select ONE.) A) Have the coordinator emit five Task tool calls in a single response, one per city, so the subagents run in parallel. B) Use fork_session five times from the coordinator's session, so each branch researches one city while sharing the full history. C) Spawn one subagent with all five cities in its prompt, so a single delegation covers the whole set of cities at once. D) Keep sequential delegation but lower each search subagent's max_tokens, so each of the five delegations finishes more quickly.

Scenario 4 · Structured Data Extraction

D3 · Prompting & Structured Output

Q574. Fenwick now checks 80-page carrier rate agreements in a single request that compares every clause with invoice terms. Results are uneven: early clauses get detailed checks and later ones superficial notes, and the same fuel clause is approved in one section but flagged in another. How should you restructure the review? (Select ONE.) A) Use a model with a larger context window, so that the whole agreement fits comfortably and attention spreads evenly over all 80 pages. B) Move the instructions to the end of the prompt, after the agreement, so the model remembers the criteria when it reaches later clauses. C) Ask the model to review the agreement twice in the same request, and keep only the issues that both passes of the review agree on. D) Run focused passes over each section for local issues, then a separate integration pass that compares findings across sections for consistency.

D3 · Prompting & Structured Output

Q575. Fenwick's contract says every backlog document must be extracted within 36 hours of arrival. Documents arrive around the clock, and the backlog runs on the Message Batches API, which can take up to 24 hours to finish a batch. How often must the team submit batches to guarantee the 36-hour limit? (Select ONE.) A) At least every 12 hours, because a document waits at most 12 hours for submission plus up to 24 hours of processing. B) Whenever 10,000 documents have accumulated, since large batches process most efficiently and arrival volume is fairly steady. C) Every 36 hours, matching the contract window, so each batch contains exactly one contract period's worth of new documents. D) Once a day at a fixed time, since most batches finish in far less than 24 hours and the remaining margin covers the rest.

D5 · Context & Reliability

Q576. Fenwick's extraction reaches 99.3% field accuracy overall on last month's audited sample, and management wants to stop human review of high-confidence output. Scanned fuel-surcharge notices are 6% of volume and barely appeared in that sample. What should happen before review is reduced? (Select ONE.) A) Reduce review now, since 99.3% overall already exceeds the client's 99% requirement and confidence filtering catches the rare errors left. B) Raise the confidence threshold for skipping review from 0.90 to 0.97, so only the most confident extractions bypass the auditors. C) Measure accuracy by document type and field on a labeled set with enough scanned notices, and keep stratified sampling of high-confidence output afterwards. D) Keep full human review of every document indefinitely, since any automation risks missing errors that auditors would otherwise catch.

D3 · Prompting & Structured Output

Q577. A Fenwick batch of 25,000 documents completes with 310 failed requests. Most failures are 140-page bills of lading that exceeded the context limit. Each request carried a custom_id set to its document ID. What should the team do next? (Select ONE.) A) Resubmit the 310 failed documents unchanged, since batch failures are usually transient and a second attempt will probably succeed. B) Use the custom_id values to resubmit only the 310 failed documents, splitting the oversized bills of lading into chunks before resubmission. C) Resubmit the entire batch of 25,000 documents, since partial results could be inconsistent with a later separate rerun of only the 310 failed ones. D) Move the 310 failed documents to the synchronous API with the same full content, since synchronous requests accept larger inputs.

D3 · Prompting & Structured Output

Q578. Fenwick's flag output covers five categories. Auditors confirm 90% of 'duplicate invoice' and 'rate above contract' flags but only 8% of 'unusual accessorial charge' flags, and they now distrust the whole flag list. You need time to rework that category's criteria. What should you do now? (Select ONE.) A) Lower the reporting threshold for all five categories at once, so that the model reports fewer flags overall across every category. B) Keep all five categories, but sort each invoice's flags so the high-precision categories always appear first and the noisy one appears last in the list. C) Temporarily stop emitting 'unusual accessorial charge' flags to restore trust in the accurate categories, while you rewrite and test its criteria. D) Add a disclaimer to every 'unusual accessorial charge' flag saying it may be unreliable, and keep sending it to auditors while the prompt is reworked.

D1 · Agentic Architecture

Q579. An engineer proposes replacing Fenwick's processing with an autonomous agent that decides at runtime which steps to run for each document. Today every document goes through the same steps (classify, extract, validate totals, normalize units, load), and the order never changes. Which design should you keep or adopt? (Select ONE.) A) Merge all five steps into one large prompt, so the model handles classification, extraction, validation and loading in a single call. B) Run the five steps in parallel for each document, since they are independent tasks, and merge their outputs before the load step. C) Adopt the autonomous agent, since letting the model choose steps per document adapts better to the wide variety of carrier layouts. D) Keep a fixed sequential prompt chain with checks between stages, and reserve dynamic decomposition for open-ended investigation tasks.

D3 · Prompting & Structured Output

Q580. Besides extracting fields, Fenwick asks Claude to flag 'anything suspicious' on each invoice for auditors, adding 'only flag high-confidence issues.' Auditors dismiss 62% of the flags, mostly normal fuel surcharges and rounding differences under $1, and have started ignoring flags altogether. What change best improves precision? (Select ONE.) A) Replace the vague request with explicit categories to flag, such as duplicate invoice numbers, and list what to skip, such as rounding under $1. B) Have a second Claude call review every flag and remove any it considers doubtful before the flags reach the auditors' queue. C) Strengthen the instruction to 'be very conservative and flag only issues you are certain about', and repeat it at both the start and end of the prompt. D) Ask Claude to attach a confidence score to each flag, and show auditors only the flags scored above 0.9 to bring the dismissal rate down.

D3 · Prompting & Structured Output

Q581. Fenwick classifies each document as invoice, bill_of_lading or surcharge_notice before extraction. Carriers increasingly send combined documents, such as a bill of lading with a charges table attached, and classification of these varies from run to run. Prose rules for combined documents have not helped. What should you add? (Select ONE.) A) Few-shot examples of combined documents showing the chosen type and a short reason it beats the plausible alternative, so the model generalizes. B) A new enum value for every combined layout seen so far, so that each carrier's combined document receives its own exact and stable classification. C) A higher temperature for the classification call, followed by a majority vote across five runs to choose the final document type. D) A rule in code that classifies any document containing a charges table as an invoice, applied before the model sees the document.

D5 · Context & Reliability

Q582. Each month, Fenwick asks Claude to write a carrier performance summary from 40 per-carrier extraction reports joined into one long input. Auditors notice that problems for carriers in the middle of the input are often missing from the summary, while the first and last carriers are covered well. What change best addresses this? (Select ONE.) A) Ask the model in the prompt to read the whole input carefully and to give equal attention to every one of the 40 carriers in the reports. B) Sort the 40 reports alphabetically by carrier name, so the order is predictable and auditors can check the summary's coverage more easily. C) Put a key-findings summary of all carriers at the start of the input, and give each detailed report its own section header. D) Increase max_tokens for the summary request, so the model has room to write about every carrier instead of cutting the summary short.

D1 · Agentic Architecture

Q583. Fenwick's dispute agent calls lookup_carrier_contract and extract_invoice_fields as needed. To save tokens, the loop sends each new request with only the system prompt, the original task and the latest tool result. Traces show the agent requesting the same contract lookup four or five times per dispute. What should the loop do? (Select ONE.) A) Add a system prompt rule telling the agent never to call lookup_carrier_contract more than once for any single dispute it handles in a session. B) Append each assistant response with its tool_use blocks, then the matching tool_result blocks, to the messages array sent on every request. C) Force tool_choice to extract_invoice_fields after the first contract lookup, so that the agent cannot request the same contract lookup again. D) Cache lookup_carrier_contract results in the tool server, so repeated calls for the same contract return instantly without another database query.

D3 · Prompting & Structured Output

Q584. About 900 fuel-surcharge notices a month fail validation because contract_rate is empty. Investigation shows these notices only say 'rates per your master agreement', a separate document that is not included in the request. The team plans to add three retries with error feedback. What should you do instead? (Select ONE.) A) Add few-shot examples of surcharge notices with contract_rate filled in, so that the model learns where the rate usually appears on notices from each of the carriers. B) Recognize that retries cannot recover information absent from the input: supply the master agreement, or return contract_rate as null for review. C) Make contract_rate required and add an enum of the most common contract rates, so that the model picks the most likely rate for each notice it processes. D) Add the retries with error feedback as planned, including the exact missing-field message, so the model knows contract_rate must be filled in.

D3 · Prompting & Structured Output

Q585. Fenwick's validator checks that line items sum to the invoice total. About 4% of extractions fail this check, usually because a credit line was read as a charge. Failed documents are simply re-sent with the same prompt, and most fail again. How should the retry work? (Select ONE.) A) Send a follow-up with the original document, the failed extraction and the specific validation error, and ask for a corrected extraction. B) Route every document that fails the sum check directly to human auditors, since another model attempt cannot fix an arithmetic error like this one. C) Switch the extraction to strict tool use with the JSON schema, which guarantees that the extracted line items always sum to the stated invoice total. D) Re-send the same prompt up to five times and keep the first extraction that passes the sum check, since model outputs vary from run to run.

D4 · Tools & MCP

Q586. Fenwick exposes one generic process_document tool whose input is free text and whose output changes by mode: extract fields, verify a total, or summarize. Traces show Claude often runs it in the wrong mode, and downstream code breaks on the varied output shapes. What change fits best? (Select ONE.) A) Keep process_document and add few-shot examples of correct mode choices to the system prompt, leaving the tool definition unchanged. B) Split it into purpose-specific tools such as extract_invoice_fields, verify_invoice_total and summarize_document, each with its own description and schemas. C) Add a post-processing layer in code that detects which output shape came back and converts it into the single fixed format that the downstream code expects. D) Keep process_document but add a required mode parameter with an enum of the three modes, and describe each of the modes in detail in the system prompt.

D3 · Prompting & Structured Output

Q587. Fenwick's extraction returns null for fuel_surcharge on 14% of documents that actually contain one. Reviews show the value appears in varied forms: a line item, a footnote like 'FSC 11.5% applied', or a percentage inside the rate table. The prompt already describes all three forms in prose. What is the most effective next step? (Select ONE.) A) Rewrite the prose to describe the three forms in more detail, with exact wording for each, and move that description to the end of the prompt. B) Run a second extraction call for fuel_surcharge only, and keep the second answer whenever the first call returned a null value for that field. C) Add two to four few-shot examples showing documents with each surcharge form and the correct extracted value, plus one with no surcharge. D) Make fuel_surcharge a required field in the schema, so that the model must return a value for every document instead of leaving the field null.

[
  {"document_excerpt": "Line 4: Fuel surcharge ..... 245.10", "fuel_surcharge": {"amount": 245.10, "percent": null}},
  {"document_excerpt": "* FSC 11.5% applied to linehaul", "fuel_surcharge": {"amount": null, "percent": 11.5}},
  {"document_excerpt": "Rate table: Linehaul 2,130.00 | FSC 9.8% | Total 2,338.74", "fuel_surcharge": {"amount": null, "percent": 9.8}},
  {"document_excerpt": "Linehaul 1,450.00 | Detention 75.00 | Total 1,525.00", "fuel_surcharge": null}
]

D3 · Prompting & Structured Output

Q588. Fenwick's extract_invoice schema marks every field required, and accessorial_type is an enum of six charge types. Audits find invented PO numbers on invoices that have none, and new charges such as 'lift-gate service' forced into the closest enum value. Which two schema changes address these problems? (Select TWO.) A) Make po_number and other fields that documents may omit optional or nullable, so the model can return null instead of inventing a value. B) Add an 'other' value to the accessorial_type enum with an accompanying detail string, so new charge types are captured rather than misfiled. C) Add a description to every required field telling the model never to invent values, while keeping all of the fields required. D) Replace the accessorial_type enum with a free-text string field, so the model can write any charge type it finds on the document.

{
  "type": "object",
  "properties": {
    "po_number": {"type": ["string", "null"]},
    "accessorial_type": {"type": "string", "enum": ["detention", "fuel", "residential", "inside_delivery", "redelivery", "other"]},
    "accessorial_detail": {"type": ["string", "null"], "description": "Required when accessorial_type is other: the charge as written."}
  },
  "required": ["po_number", "accessorial_type", "accessorial_detail"]
}