Concept Drills — 120 Questions That Test the Concept, Not the Shape

Exam items rarely reuse a scenario you have seen; they reuse the concept behind it. Each drill here tests one concept from the exam blueprint through a fresh scenario, a different angle and its own options, so recognising a question's shape stops being enough. Name the concept first, then choose. Where a question raised a doubt during review, the resolved clarification sits beneath it.

Jump to a domain: Agentic Architecture · Claude Code · Prompting & Structured Output · Tools & MCP · Context & Reliability

These drills share one grading file with the other practice pages; resetting a domain here never touches another page's score.

Drill · Agentic Architecture (Q220–Q249)

Q220. A freight company's coordinator asks a subagent to normalise 40,000 shipping addresses. The delegation prompt is eleven numbered steps (split on commas, uppercase the third field, pad postcodes to five digits, and so on). Partway through, the subagent meets a new region whose addresses contain no commas, keeps executing step two anyway, and reports success on thousands of mangled rows. Which diagnosis is exactly right? A) The subagent had too many tools available, which distracted it from the address-parsing steps it was supposed to follow in order. B) The subagent ran out of output budget while reasoning about the new format, so it fell back to the literal steps it had been given. C) The steps were sound; the prompt only needed a rule telling the subagent to halt and ask the coordinator whenever input deviates from the script. D) The prompt specified how to work but never what a correct result looks like, so when the data broke the script the subagent had no criterion to notice failure or adapt toward.

Q221. A documentation platform's coordinator receives two jobs at once: correct a misspelled word in one README heading, and audit 400 API reference pages for broken cross-links. The coordinator already has the README open. Which assignment keeps coordination cost lowest while keeping the coordinator's own context clean? A) Two subagents, one per job, so the coordinator stays a pure router and never edits files itself. B) A subagent fixes the heading, while the coordinator audits the pages because link checking needs the global view only it holds. C) The coordinator fixes the heading itself and hands the 400-page audit to one subagent that returns only the list of broken links. D) The coordinator does both, reading all 400 pages into its own context so it has a complete picture for future requests.

Q222. A market-intelligence system is asked to profile the payment methods used by consumers across Southeast Asia. The final report covers credit cards and bank transfers in depth but never mentions e-wallets, buy-now-pay-later or cash on delivery. Reviewers agree each worker's section is thorough and well sourced. Which single piece of evidence would most directly confirm where the gap originated? A) The synthesis worker's prompt, to check whether it was told to flag categories missing from its inputs. B) The retrieval index, to verify whether documents about e-wallets and buy-now-pay-later exist in the corpus at all. C) The coordinator's delegation record showing the list of subtopics it assigned before any worker started. D) The search worker's query log, to see whether its queries were phrased too narrowly to surface wallet providers.

Q223. A support agent's loop ends when the assistant's text contains the sentinel "FINAL ANSWER:". Some tickets end mid-task because the agent quotes a customer email that happens to contain that phrase; others grind on until the 25-turn safety cap even though the work finished earlier. Which statement about the loop's exit condition is exactly right? A) A rarer sentinel such as a random token would fix both failures, because customers could never type it by accident. B) Every response already carries a stop_reason that separates wanting a tool (tool_use) from being finished (end_turn), so the loop should branch on that field rather than on text. C) Keep the sentinel but normalise case and require it at the very end of the message, so quoted text cannot trigger it. D) The 25-turn cap is the real defect; raising it lets long tickets reach the sentinel before the loop gives up.

Q224. A travel-planning coordinator runs a flight agent and then a hotel agent. The flight agent finds that the only affordable flight lands a day later than the traveller asked for and returns that itinerary. The hotel agent then books rooms starting on the originally requested date, leaving a paid night unused. The two agents never exchange messages. Which change fixes this at the right layer? A) Add a shared table that both agents poll while they work, so each can see the other's latest decisions. B) Run the hotel agent first, so the booked dates constrain which flights the flight agent is allowed to choose. C) The coordinator puts the flight agent's confirmed arrival date and time into the hotel agent's delegation prompt. D) Give the hotel agent read access to the flight agent's working log so it can find the arrival date on its own.

{
 "task": "Book a hotel near the conference venue",
 "constraints": {
  "check_in": "2026-11-04",
  "check_out": "2026-11-07",
  "max_nightly_rate_usd": 220
 },
 "from_flight_agent": {
  "arrival_local": "2026-11-04T21:40",
  "airport": "SIN",
  "note": "Only affordable option lands one day after the requested date"
 },
 "return_format": "hotel name, confirmation id, total cost"
}

Q225. A hospital's procurement officer wants three infusion-pump vendors compared on recall history, five-year total cost of ownership and nurse usability, with every criterion measured the same way for all three vendors so the committee can read the table across. Which decomposition is MOST likely to produce a comparison that cannot be read side by side? A) One worker per criterion, each covering all three vendors, followed by a synthesis step. B) One worker per criterion, with the coordinator fixing the metric definition in each brief, then synthesis. C) One worker per vendor, each researching all three criteria for its own vendor, merged at the end. D) A single worker researching every vendor and criterion in one pass, filling a supplied comparison template.

Q226. A platform team is choosing an orchestration style for four upcoming jobs. Which TWO jobs are better served by a fixed prompt chain than by dynamic adaptive decomposition? (Select TWO.) A) Migrating an undocumented internal library whose callers and hidden dependencies have not yet been mapped. B) A weekly compliance check that runs the same five named checks on every pull request in a fixed order. C) Investigating why a mobile app's crash rate doubled after an unknown combination of recent releases. D) Nightly processing of supplier invoices: extract fields, validate totals, post to the ledger, the same three steps every night.

Q227. An analytics team inherits a data warehouse of 900 tables with no lineage documentation, and nobody knows which tables actually feed the revenue dashboards. An agent is asked to produce useful documentation for the warehouse. What should the agent's first phase be? A) Trace lineage backward from the dashboards' queries to find which tables matter, then set documentation depth according to what that map reveals. B) Document every table alphabetically with one uniform template so coverage is even and progress is easy to track. C) Write the complete documentation outline for all 900 tables in a single pass before looking at any query history. D) Split the 900 tables evenly across ten parallel workers, each documenting its 90 tables in full.

Q228. A law firm's intake assistant has five specialists: conflict checking, jurisdiction research, fee estimation, document drafting and translation. Its coordinator invokes all five for every inquiry, including "What are your office hours?", and the partners complain about cost and slow replies. Where should the fix live? A) In the coordinator: assess each inquiry and dispatch only the specialists it actually needs. B) In the front end: add a separate FAQ bot so simple questions never reach the multi-agent system. C) In the specialists' configuration: lower each one's output budget so unnecessary runs finish faster. D) In each specialist: add an early check that returns "not applicable" when an inquiry is outside its area.

Q229. A security analyst has spent a long session triaging an intrusion and building a detailed attack timeline. She now wants to evaluate two containment plans (isolate the affected subnet, or rotate every credential and keep services running), each reasoned from the same full timeline, without either plan's reasoning leaking into the other. What should she do? A) Export the timeline to a notes file and open two new sessions that each read the file. B) Resume the session, evaluate the subnet plan, then clear the conversation and evaluate the credential plan. C) Fork the session from the finished timeline twice and evaluate one containment plan in each branch. D) Ask the agent to evaluate both plans side by side in a single reply so neither one is favoured.

# baseline session id comes from the triage run
claude --resume 7c1e-triage --fork-session -p "Evaluate containment plan one: isolate the affected subnet."
claude --resume 7c1e-triage --fork-session -p "Evaluate containment plan two: rotate all credentials, keep services running."

Q230. During a long session tuning a pricing service, operations roll back a deployment and several configuration files on disk change. The agent keeps quoting the old values it read an hour ago. A teammate says, "Just tell it to reread the changed files." Why is that alone unlikely to fully fix the problem? A) The context window is already full, so any new file contents are silently dropped before the model sees them. B) Rereading works fully; the agent simply needs extended thinking enabled to notice that the values changed. C) The earlier reads stay in the conversation beside the new ones, so the agent can still anchor on the outdated values in later reasoning. D) File tools cache results for the lifetime of a session, so a reread returns the same stale content.

Q231. A due-diligence system produces a memo on an acquisition target. The coordinator's review finds that one revenue figure has no source and the question about pending litigation was never answered; everything else is solid. Which next step keeps the good work and closes the holes? A) Ask the synthesis worker to supply the revenue source and litigation status from what it already knows. B) Send narrowly scoped follow-up tasks for the unsourced revenue figure and the litigation question, then rerun synthesis with the new findings. C) Discard the memo and rerun every worker with a longer, more detailed brief. D) Mark the two gaps as unverified and deliver the memo, letting the client's analysts chase the missing evidence.

Q232. An insurance intake agent is meant only to collect and verify policy details, but it has eleven tools, including approve_payout, draft_denial_letter and score_fraud_risk. In testing it sometimes drafts denial letters during intake. A system-prompt warning has already been tried. Which fix removes the cause rather than the symptom? A) Give the intake agent only its collection and verification tools, and move payout, denial and fraud tools to the adjudication agent. B) Strengthen the system prompt with a bold rule that the intake agent must never use adjudication tools. C) Keep all eleven tools but add a hook that rejects adjudication tool calls coming from the intake agent. D) Rewrite the adjudication tools' descriptions to say they are only for use after intake is complete.

Q233. A release pipeline uses three generator subagents: one writes the iOS client, one the Android client, and one the shared API contract. After a run, Android sends timestamps as epoch milliseconds, iOS sends ISO strings, and the contract expects ISO strings. Each generator's output is internally consistent. Where should the consistency check live? A) In the coordinator's own planning turn, with extended thinking enabled so it can compare the three outputs itself. B) In a dedicated reviewer that receives all three outputs from the coordinator and checks only cross-component agreement. C) In each generator, by letting it browse the other generators' working directories while it runs. D) In each generator, re-prompted with an instruction to match the other platforms' formats before returning.

Q234. A database team's coordinator runs a migration-writer subagent and then a rollback-writer subagent for each ticket. Reviewers notice the rollback scripts reliably undo what each ticket describes, but miss extra changes the migration writer actually made, such as an index it added for performance. What explains the pattern? A) The tickets are too vague; more precise tickets would keep both scripts in agreement. B) The rollback writer lacks a database connection to inspect the live schema after the migration runs. C) The two writers run on different model tiers and therefore interpret the ticket differently. D) The rollback writer receives only the ticket, never the migration script the first writer produced.

Q235. An IT help-desk agent runs a fourteen-question diagnostic interview (operating system, network, recent updates, log excerpts) on every ticket, including "Please reset my password". Staff rate the assistant as exhausting, yet for real outages the full interview is what engineers need. Which change serves both kinds of ticket? A) Triage each ticket by type first and pick the interview depth from that type: none for a password reset, the full diagnostic for an outage. B) Shorten the wording of each of the fourteen questions so the interview is quicker to answer. C) Drop the interview and require staff to complete a detailed form before any ticket is accepted. D) Cap the interview at four questions for every ticket.

Q236. An agent must write a SQL query whose output matches a finance team's reference report. On each retry it writes a brand-new query from scratch, and after eight attempts the mismatch persists in different places each time. What input is the retry loop missing? A) A system-prompt reminder to fix the previous attempt rather than start over. B) A higher retry ceiling so the agent has more attempts to land on a matching query. C) A lower temperature so each fresh attempt is more deterministic. D) The previous query together with the specific rows where its output differed from the reference.

Q237. An e-commerce operations assistant spawns inventory, pricing and fraud specialists for every request. Asking "How many units of the blue kettle are in stock?" takes as long and costs as much as "Why did our margin drop last week?", which genuinely needs all three. Which statement about the redesign is exactly right? A) The cheapest fix is to keep all three specialists on every request but have them share one cached context. B) Because the margin investigation needs all three specialists, the stock lookup must also run them to keep output formats consistent. C) The coordinator should improvise a fresh architecture for each request, since no single policy can cover both kinds. D) The selection rule is designed once into the coordinator, and each request is assessed so the rule decides which specialists that request actually runs.

Q238. A localisation coordinator reviews four independent translation files (French, German, Japanese and Portuguese), delegating one per turn and waiting for each reviewer to finish before starting the next. Each review takes about 40 seconds, so the full pass takes nearly three minutes. What should the coordinator do differently in its output? A) Raise each reviewer's max_tokens so it can finish in a single turn. B) Emit all four review delegations as tool_use blocks in the same assistant response, so a runtime that executes them concurrently runs them together. C) Keep one delegation per turn but shorten each reviewer's prompt so it responds faster. D) Merge the four languages into one reviewer so there is only one handoff.

{
 "role": "assistant",
 "content": [
  {
   "type": "text",
   "text": "Reviewing all four locales in parallel."
  },
  {
   "type": "tool_use",
   "id": "toolu_fr",
   "name": "Agent",
   "input": {
    "description": "Review French strings",
    "prompt": "Review locales/fr.json against the glossary; return issues as a list."
   }
  },
  {
   "type": "tool_use",
   "id": "toolu_de",
   "name": "Agent",
   "input": {
    "description": "Review German strings",
    "prompt": "Review locales/de.json against the glossary; return issues as a list."
   }
  },
  {
   "type": "tool_use",
   "id": "toolu_ja",
   "name": "Agent",
   "input": {
    "description": "Review Japanese strings",
    "prompt": "Review locales/ja.json against the glossary; return issues as a list."
   }
  },
  {
   "type": "tool_use",
   "id": "toolu_pt",
   "name": "Agent",
   "input": {
    "description": "Review Portuguese strings",
    "prompt": "Review locales/pt.json against the glossary; return issues as a list."
   }
  }
 ]
}

Q239. A team asks an agent, in one instruction, to internationalise a 25-screen app: find every hard-coded string, replace each with a translation key, and reconcile duplicate keys across screens. The result is inconsistent: some screens are clean, others keep hard-coded text, and the same phrase gets different keys on different screens. Which explanation best accounts for this? A) The prompt lacked examples; two well-chosen examples in the single instruction would make every screen consistent. B) The response was truncated, so a larger max_tokens would have let the agent finish consistently. C) The model cannot edit more than a few files per request, so the work should be split only by file count. D) One instruction asked for three different jobs across 25 screens at once, so attention per job thinned; each stage needs its own focused pass.

Q240. A product analyst spent an hour in an agent session building a churn analysis of subscription cohorts. She now wants to test two pricing ideas, an annual-plan discount and a usage-based tier, each starting from that analysis, without either exploration influencing the other. Which statement about using a session fork for this is correct? A) A fork keeps only a compacted summary of the original session, trading detail for isolation. B) A fork is equivalent to resuming; both simply append new turns to the original session's history. C) A fork creates a new session whose history begins as a copy of the original at the fork point, so each branch keeps the full analysis and neither sees the other's later turns. D) A fork shares one live history, so turns in either branch appear in both and keep them in sync.

from claude_agent_sdk import query, ClaudeAgentOptions

async def explore(baseline_id, idea):
    opts = ClaudeAgentOptions(resume=baseline_id, fork_session=True)
    async for msg in query(prompt=f"Using the churn analysis so far, evaluate: {idea}", options=opts):
        print(msg)

# await explore(saved_id, 'annual-plan discount')
# await explore(saved_id, 'usage-based tier')

Q241. A pre-launch readiness review spawns legal, accessibility and performance reviewers in parallel. All three return successfully and no errors are logged, yet the final readiness report contains only accessibility findings. Which control would have caught this before the report went out? A) Before synthesis, the coordinator checks that it holds a result for every reviewer it dispatched and passes all of them into the synthesis step. B) Tell the synthesis step to write "no issues found" for any area without findings. C) Automatically retry any reviewer whose call returns an error. D) Raise each reviewer's turn limit so it has room to finish its analysis.

Q242. A compliance agent scans 60 vendor contracts for a nonstandard indemnity clause and runs with a 10-turn limit. In about half the runs it stops after roughly 35 contracts, missing later matches; in the rest it reaches all 60. No tool errors are logged. What is the root cause? A) Contracts after the 35th are longer, so the model stops reading them in detail. B) The system prompt does not list all 60 contract file names, so the model loses track of which contracts remain. C) The turn limit is set below what the task needs, so the loop is cut off mid-scan; completion should come from end_turn, with any cap kept as a generous safety ceiling. D) A rate limit silently drops tool calls after about 35 contracts.

Q243. A team configures an incident-postmortem orchestrator with the Agent SDK. Its system prompt says to hand timeline reconstruction and log analysis to the subagents defined in its configuration. Every run finishes in a single turn with the orchestrator writing a shallow postmortem itself; no subagent ever starts and no errors appear. Which configuration problem is MOST likely? A) The permission mode is set to accept edits, which disables all subagents. B) The delegation tool (named Agent in current SDK docs, Task in older material) is not among the tools available to the orchestrator, so it has nothing to call. C) The subagent definitions omit a model field, so they cannot start. D) The system prompt says "hand off" instead of naming each subagent's identifier.

Q244. An inventory-reconciliation agent's loop returns as soon as a response contains any text block. When the model writes "Let me check the warehouse count" and also requests a tool in the same response, the loop exits before the tool ever runs. Which rule should the loop follow instead? A) Return when the text mentions a final reconciled count; otherwise keep looping. B) If stop_reason is tool_use, run every requested tool, append the tool_result blocks in a user message and call again; if it is end_turn, return the answer. C) Return when the response contains any text; otherwise execute the tools. D) Run exactly five iterations and return whatever text the last response contains.

def run(client, model, messages, tools, execute):
    while True:
        resp = client.messages.create(model=model, max_tokens=2048, tools=tools, messages=messages)
        messages.append({"role": "assistant", "content": resp.content})
        if resp.stop_reason == "tool_use":
            results = [
                {"type": "tool_result", "tool_use_id": b.id, "content": execute(b.name, b.input)}
                for b in resp.content if b.type == "tool_use"
            ]
            messages.append({"role": "user", "content": results})
            continue
        if resp.stop_reason == "end_turn":
            return resp
        raise RuntimeError(f"unhandled stop_reason: {resp.stop_reason}")

import os
from typing import Annotated

import anthropic
from anthropic import beta_tool
from pydantic import Field

from app.repo import code_search  # your own module

client = anthropic.Anthropic()  # reads ANTHROPIC_API_KEY
MODEL = os.environ.get("CLAUDE_MODEL", "claude-sonnet-5")


@beta_tool(
    name="search_code",
    description="Search the repository and return matching paths with line numbers.",
)
def search_code(pattern: Annotated[str, Field(description="Text or a regular expression to find, e.g. legacyAuthenticate.")]) -> str:
    return code_search.grep(pattern)


def run(task: str):
    runner = client.beta.messages.tool_runner(
        model=MODEL, max_tokens=2048, tools=[search_code],
        messages=[{"role": "user", "content": task}],
        max_iterations=20,  # a safety ceiling, not the success signal
    )
    last = None
    for last in runner:  # one assistant message per turn; narration text does not stop it
        print(last.stop_reason, last.usage.output_tokens)
    if last is not None and last.stop_reason == "end_turn":
        return last
    raise RuntimeError(f"stopped without end_turn: {last and last.stop_reason}")

Q245. A property-valuation assistant asks a comparable-sales subagent (about 2.4 s) and then a zoning-rules subagent (about 2.1 s), one after the other, for each address. End-to-end latency is about 5 s, and the two lookups are independent. Which change lowers latency the most, and to roughly what? A) Request both subagents in the same coordinator response so they overlap, bringing latency down toward the slower lookup, around 2.5 to 3 s. B) Put a 1-second timeout on the zoning subagent so the sales subagent can start sooner. C) Switch the coordinator to a smaller, faster model, because its planning turns dominate the 5 seconds. D) Fold both lookups into the coordinator's own context to avoid the cost of spawning subagents.

Q246. A curriculum builder uses research subagents to gather material, then drafts a course outline in a single pass. On broad subjects such as "climate science for policymakers", some modules end up thin or empty, but the coordinator declares the outline finished anyway. Which coordination pattern most directly improves coverage? A) Have the final outline list thin modules in an appendix so readers know where to look elsewhere. B) Give each research subagent a larger output budget so it returns more material per call. C) Spawn twice as many research subagents at the start of every run. D) Have the coordinator check the draft against the intended coverage, send targeted research tasks for thin modules, and repeat until every module meets the bar.

Q247. A meal-planning service runs a nutrition subagent and an allergy subagent, then spawns a plan-writer subagent. Both analysis subagents succeed, but the meal plans reflect only the nutrition findings, and some include foods the allergy check flagged. Logs show the plan-writer's prompt contained the nutrition output alone. Which statement is exactly right? A) Subagents automatically inherit the coordinator's full history, so the allergy output must have been truncated by the plan writer. B) The plan writer knows only what its delegation prompt contains or explicitly references, so the allergy findings never reached it. C) The allergy results were returned as structured JSON, which subagent prompts cannot accept. D) Subagents can read each other's results through a shared session store unless isolation is turned off.

Q248. A policy institute asks its research system about the consequences of a four-day work week. The coordinator's plan lists four subtasks, all about employee outcomes (productivity, burnout, retention, commuting), and the report never covers employer costs, customer-service coverage or effects on public services. Which change prevents this class of failure? A) Have the coordinator check its subtask list against the full scope of the question, adding missing dimensions before any delegation. B) Instruct search workers to broaden their queries beyond the subtopic they were assigned. C) Tell the synthesis worker to add sections for obvious omissions from its own general knowledge. D) Loosen the analysis worker's relevance filter so more borderline sources pass through.

Q249. A medical-billing pipeline runs three steps: identify the claim type, assign billing codes, then check that codes agree with the diagnosis. Dental claims reach 71% accuracy while outpatient claims reach 96%. An engineering lead proposes merging all three steps into one prompt to cut cost by more than half. What evidence would need to exist before the merge is acceptable? A) A pilot running the merged prompt on outpatient claims only, since they are already accurate. B) An evaluation on labelled claims showing the merged prompt matches the pipeline's accuracy for each claim type, dental included, and still catches code-diagnosis mismatches. C) A latency benchmark showing the merged prompt is faster on average. D) Sign-off from the product owner accepting somewhat lower accuracy on dental claims.

Drill · Claude Code (Q250–Q274)

Q250. A legal-research group runs Claude Code sessions that last most of a working day. Auto-compaction regularly kicks in, and afterwards the case citations Claude gathered early on are gone from the conversation. The group wants a script that pulls every citation out of the transcript at the exact point before anything is summarised, whether a person typed the command or the context limit triggered it. On which hook event should the script be registered? A) Stop, because it fires each time Claude finishes a response and can therefore scan the transcript frequently B) SessionEnd, because it fires once the session closes and can read the complete transcript in one pass C) UserPromptSubmit, filtering for prompts whose text starts with the compaction slash command D) PreCompact with no matcher, so it runs before both manual and automatic compaction and receives the transcript path

{
  "hooks": {
    "PreCompact": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "python3 .claude/hooks/extract_citations.py"
          }
        ]
      }
    ]
  }
}

Q251. A data-science platform requires every Python file to use a 100-character line limit, and the build fails on anything else. The rule sits in the committed project CLAUDE.md. A contractor's personal instruction file says "wrap at 79 characters", and several of the contractor's sessions have produced files that failed the build. Which statement about fixing this is exactly right? A) Adding a lineLength key to the committed .claude/settings.json makes Claude Code enforce the width on every generated file B) Any CLAUDE.md level is guidance the model weighs; a PostToolUse hook on Write and Edit that runs the formatter makes the width deterministic C) Project CLAUDE.md is loaded after the personal file, so it deterministically overrides it and the failures must come from something else D) Moving the rule into an organisation-wide managed CLAUDE.md guarantees compliance because users cannot override managed instructions

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Write|Edit",
        "hooks": [
          {
            "type": "command",
            "command": "jq -r '.tool_input.file_path' | xargs ruff format --line-length 100"
          }
        ]
      }
    ]
  }
}

Q252. A mobile team's CLAUDE.md says "new screens must support accessibility." Claude Code keeps producing screens where a single icon has a content description, while buttons have no labels and touch targets are too small. Tests and CI are already configured. What is the root cause of the thin results? A) The instruction names a goal but no observable criteria, so Claude meets it minimally; it needs checkable requirements that tests can verify B) The instruction must live in a .claude/rules file to become binding, since CLAUDE.md content is treated as optional background C) CLAUDE.md is only consulted for files at the repository root, so screen files in nested folders never see the instruction D) Claude Code drops CLAUDE.md after each compaction, so the instruction fades as the session grows longer

Q253. A game studio keeps separate standards for shader code, level data files and localisation strings. A lead put each standard in its own file under .claude/rules/ so that only the relevant one would be in context while Claude works in that area. None of the files has frontmatter. What actually happens? A) All three files load at the start of every session, so every standard is in context regardless of what is edited B) Claude Code infers the scope from each file name and loads it only for folders with a matching name C) Each file stays dormant until a CLAUDE.md imports it, so none of the three standards reaches Claude D) Claude Code refuses to start and reports that every rules file requires a paths header

---
paths:
  - "shaders/**"
  - "**/*.hlsl"
---
# Shader standards
- Use half precision for colour values unless a comment justifies float.
- Every shader declares its target profile at the top.

Q254. A Go team registers gofmt -w on the edited file as a PreToolUse hook matching Edit and Write, expecting every file Claude touches to end up formatted. After a week, files Claude modified are still unformatted in review. What is the cause? A) Formatting is only applied when the requirement is written in CLAUDE.md so that Claude runs the tool itself B) Hooks cannot launch external binaries, so gofmt is skipped and only built-in checks run C) The matcher must be Bash because formatters are shell commands, so an Edit matcher never fires D) PreToolUse runs before the edit is applied, so gofmt formats the old file and Claude's change lands afterwards

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Edit|Write",
        "hooks": [
          {
            "type": "command",
            "command": "jq -r '.tool_input.file_path' | xargs gofmt -w"
          }
        ]
      }
    ]
  }
}

Q255. At a fintech company, two developers each start a Claude Code session in the same checkout: one upgrades the logging library, the other adds a new report endpoint on a different branch. Within an hour one session's test run picks up the other's half-written files, and a branch switch in one session changes the files the other is editing. What is the underlying cause and fix? A) Claude Code does not support concurrent sessions; run the two tasks one after the other in a single session B) The two sessions share a single context window; clear the conversation in each session before switching tasks C) Both sessions lack a lock file in settings; enable file locking so edits from one session are hidden from the other D) Both sessions share one working directory; give each its own git worktree on its own branch and run one session per worktree

git worktree add ../app-logging -b chore/logging-upgrade
git worktree add ../app-report -b feat/report-endpoint
cd ../app-logging && claude   # session one
cd ../app-report && claude    # session two

Q256. An API rename means the client docs for four language SDKs (Python, Go, Java, Ruby) must all be rewritten, and each rewrite requires reading that SDK's source in depth. Holding all four codebases in one conversation would exceed the context window, and the work should run at the same time. Which plan meets both constraints? A) Run one session over all four SDKs in turn and compact the conversation between each language B) Start four sessions in the same checkout, each switching to its own branch before it begins editing C) Create a git worktree per SDK and run one Claude Code session in each, then merge the four branches D) Create one skill with forked context that walks through all four SDKs inside a single forked run

Q257. A health-records team requires every new database migration to be named V__.sql. Claude Code usually complies, but about one migration in ten gets a name like V3__AddTable.sql, which breaks the deploy tool. A teammate proposes adding three good and three bad examples to the project CLAUDE.md. What change actually guarantees non-conforming names never land? A) Create a .claude/rules file scoped to the migrations folder that restates the naming pattern with examples B) Add a PreToolUse hook on Write that checks the file path against the pattern and exits with code 2 and a reason C) Add the six examples to CLAUDE.md with the heading IMPORTANT so the model weighs them above other guidance D) Ask a review subagent at the end of each session to list any migrations whose names do not match the pattern

import json, re, sys

event = json.load(sys.stdin)
path = event.get("tool_input", {}).get("file_path", "")
name = path.rsplit("/", 1)[-1]
if "/migrations/" in path and not re.fullmatch(r"V\d{12}__[a-z0-9_]+\.sql", name):
    print(f"Migration name {name} must match V<yyyymmddHHMM>__<snake_case>.sql", file=sys.stderr)
    sys.exit(2)
sys.exit(0)

Q258. Before changing how permissions are checked in a large Rails monolith, a developer asks Claude Code how authorization works. The logic is spread across concerns, mixins, callbacks and a policy layer in roughly thirty files. Claude opens two files and gives a confident but partial answer. Which approach gives a thorough answer without filling the main conversation with file dumps? A) Paste the full list of thirty file paths into the prompt and ask Claude to read all of them in this session B) Run the compaction command first so that the main conversation has more room for the files it reads C) Enable extended thinking so Claude reasons more carefully about the two files it has already read D) Delegate the tracing to the Explore subagent, asking it to follow every call path and return a summary

Q259. A nightly job runs claude -p with the prompt "Confirm every feature flag in flags.yaml is still referenced in the code" and a JSON output format. It reports "all flags in use" every night, yet engineers keep finding dead flags. What is the most likely cause? A) A single instance cannot check itself; a second headless instance must confirm every verdict the first one gives B) The prompt asks for a verdict with no required evidence, so Claude samples some flags and generalises; require each flag listed with the file and line where it is used C) The JSON output format truncates long answers, so the list of unused flags is cut off before it is printed D) Print mode runs without file tools, so Claude answers from the prompt text alone and never reads flags.yaml

claude -p "For EVERY key in flags.yaml, search the source tree.
Return JSON: [{\"flag\": ..., \"references\": [\"path:line\", ...]}].
A flag with an empty references list is UNUSED. Do not summarise; list every flag." --output-format json

Q260. Compliance at an insurance firm states that Claude Code must never open files under the secrets/ folder in any engineer's session. The rule is currently a sentence in the project CLAUDE.md, and one engineer's personal instruction file tells Claude to "read whatever you need to finish the task." Where should the rule live so that it holds in every session? A) In a CLAUDE.local.md at the repository root, because it is read last and overrides earlier instruction files B) In the committed .claude/settings.json as a deny rule on reading secrets/, backed by a hook for shell access C) In an organisation-wide managed CLAUDE.md, because personal instruction files cannot remove managed instructions D) In a .claude/rules file with paths set to secrets/, so the prohibition loads whenever those files come up

{
  "permissions": {
    "deny": [
      "Read(./secrets/**)",
      "Edit(./secrets/**)",
      "Bash(cat secrets/*)"
    ]
  }
}

Q261. A platform engineer wrote a release-notes skill that drafts notes from merged pull requests, and it works well on her laptop. Teammates who pulled the latest main branch report that it does not exist for them. She had saved it at ~/.claude/skills/release-notes/SKILL.md. What is the simplest change that gives it to everyone who clones this repository? A) Register the skill by name in the project settings file so that Claude Code knows to load it for all users B) Paste the skill body into the project CLAUDE.md so that it is loaded into every teammate's sessions C) Add a user-invocable line to the skill frontmatter so that it becomes visible to users other than its author D) Move it to .claude/skills/release-notes/SKILL.md inside the repository and commit that directory

Q262. A consultant opens Claude Code inside a client's repository, which is cloned under a shared workspace folder. Claude starts insisting on a changelog entry for every change, and nothing in the repository says so. The consultant suspects an instruction file somewhere above the repository. Which command answers this directly inside the running session? A) /doctor, which checks the installation and reports problems with the Claude Code setup B) /memory, which lists the instruction files loaded for the session and lets you open each one C) /clear, which empties the conversation so the session starts again without inherited instructions D) /init, which scans the repository and writes a CLAUDE.md describing the conventions it finds

Q263. A design-system repository keeps Storybook story files next to their components, as Card.stories.tsx beside Card.tsx, across about sixty folders at different depths. The team wants its story-writing conventions applied only when Claude works on story files. Which setup does that? A) An import line in the root CLAUDE.md that pulls in a separate story-conventions file at startup B) A CLAUDE.md inside src/ with the story conventions, so they load whenever Claude works under that folder C) A rules file whose paths list is ["**/.stories.tsx"], so it applies to story files at any depth D) A rules file whose paths list is ["src/.stories.tsx"], so it applies to story files under the source folder

Q264. A team lead wrote the team's commit-message convention into her own ~/.claude/CLAUDE.md, and her sessions follow it perfectly. The CI job that runs claude -p on a build runner to prepare release commits ignores the convention entirely, and so do her teammates' sessions. What explains this? A) Print mode skips every instruction file, so headless runs never see conventions from any CLAUDE.md B) User-level instructions apply only after they are imported from the project CLAUDE.md of each repository C) User-level instructions load only in interactive sessions that are started by the same account that created the file D) The file lives in her home directory on her machine, so the runner and teammates never have it; it belongs in the committed project CLAUDE.md

Q265. A dependency-audit skill lists every transitive package, its licence and its known advisories, often thousands of lines. After running it, developers find Claude drifting toward the audit instead of the feature they were building. The team adds context: fork to the skill's frontmatter. Which statement about that change is exactly right? A) It runs the skill in a separate subagent context and returns only the final result to the main conversation B) It switches the skill to a smaller model, which produces shorter output and therefore less clutter in the session C) It stops Claude from invoking the skill automatically, so the audit only runs when a user types its name D) It copies the whole current conversation into the skill so the audit can use everything discussed so far

---
name: dependency-audit
description: Audit transitive dependencies for licence and advisory issues.
context: fork
---
List every transitive dependency with licence and known advisories.
Return only: packages needing action, why, and the suggested fix.

Q266. A developer must add account deletion to a SaaS product. Legal rules about what must be kept for audit, how backups are handled, and how long invoices are retained are not written down anywhere; only the developer and a lawyer know them. Before any code is written, what is the best way to bring the unstated constraints into the design? A) Ask Claude to scan the existing deletion jobs and infer the legal retention rules from how the code behaves today B) Start with a failing end-to-end deletion test and let each failure reveal which data and systems are involved C) Ask Claude to interview the developer with targeted questions about retention, backups and audit needs, then write a spec from the answers D) Ask Claude to compare three open-source account-deletion libraries and recommend the one with the most features

Q267. The team replaced an old logging helper and updated the project CLAUDE.md last week. Everyone's sessions now use the new helper except one developer's, which still writes calls to the old helper. That developer has pulled the latest main. The lead suspects a personal or local instruction file still describes the old helper. Which step confirms or rules out that suspicion before anything is edited? A) Make the new-helper instruction in the project CLAUDE.md more emphatic and see whether the behaviour changes B) Have the developer delete the whole ~/.claude directory so that only the project configuration remains C) Ask Claude in that session which logging helper it should use, and compare its answer with the project file D) Run /memory in that developer's session and check each loaded file for the outdated logging instruction

Q268. An incident-response team writes postmortems with twelve required sections. Engineers paste their own saved versions of a long drafting prompt into Claude Code, and the versions have drifted so some drafts miss sections. The team wants one versioned procedure anyone can run on demand with an incident ID. What should they build? A) Give each engineer a shell alias that runs Claude Code headless with the drafting prompt from a local text file B) Put the full drafting prompt in the root CLAUDE.md so it is present in every session for every engineer C) Add a rules file scoped to postmortems/** so the drafting procedure appears whenever those files are opened D) Create a committed project skill such as .claude/skills/postmortem/SKILL.md with an argument hint for the incident ID

---
name: postmortem
description: Draft a postmortem with all twelve required sections for an incident.
argument-hint: <incident-id>
---
Draft the postmortem for incident $ARGUMENTS.
Include every section: summary, impact, customer impact, timeline, detection,
root cause, contributing factors, what went well, what went poorly, actions,
owners, follow-up dates. Mark any section with missing data as TODO.

Q269. A developer has four tasks queued for Claude Code this afternoon. Which one most warrants starting in plan mode rather than asking for the change directly? A) Bump one library by a patch version exactly as a published security advisory instructs B) Move app state from a global store to a server-cache library across seventy components, with several possible migration orders C) Correct a misspelled word in one user-facing error message that a support ticket quoted exactly D) Rename a confusingly named local variable inside a single function that already has full test coverage

Q270. A monorepo's 700-line CLAUDE.md mixed Go, SQL, Helm and front-end conventions, and Go error-handling advice kept appearing in SQL files. The team split it into four files and added an import line for each at the top of the root CLAUDE.md. The bleed continues exactly as before. What is the root cause? A) Imported files are pulled in at launch, so all four sets are in context for every file; conditional loading needs rules files with paths globs B) A user-level instruction file loaded before the project file is repeating the Go advice in every session C) Imported files are loaded only after a matching file is opened, so the conventions arrive too late to shape the edit D) Import lines only resolve absolute paths, so the four files silently fail to load and older cached copies are used

Q271. A consultancy uses one repository and Claude Code to produce reports for many clients. A regulated bank needs British spelling, a mandatory risk-disclosure section and a stricter source policy, but only for its own reports; every other client keeps the defaults in the root CLAUDE.md. Analysts start each job from the client's folder. Which configuration fits? A) Add the bank's rules as environment variables in the project settings file so that they are applied to each run B) Create clients/bank/CLAUDE.md with the extra rules, so jobs started in that folder load it on top of the root defaults C) Add an "if the client is the bank" block to the root CLAUDE.md so the model applies the stricter rules when it recognises that client D) Put the bank's rules into each analyst's personal instruction file so that every report they write follows them

Q272. A database team defines three Claude Code subagents: schema-reader should only inspect code, migration-writer creates migration files, and release-runner executes the deploy script. In a trial, schema-reader edited a file while "helpfully" fixing a typo. How should the team make sure each subagent can only use the tools of its role? A) Give all three subagents every tool and have the main agent route tasks so that each role gets suitable work B) Add a deny rule for Edit and Write to the project settings so that no agent in the session can modify files C) Add a sentence to each subagent's system prompt stating which tools it may use and which it must avoid D) Set the tools field in each subagent's definition, for example Read, Grep and Glob only for schema-reader

---
name: schema-reader
description: Inspects schema and data-access code and reports findings. Never modifies files.
tools: Read, Grep, Glob
---
You analyse database schema and queries and report findings with file references.

Q273. A billing agent built on the Agent SDK can call an issue_refund tool. Finance now requires a named approver to sign off every refund before money moves; the agent should still identify refunds and prepare them. Which change meets the requirement? A) Keep issue_refund but cap each run at one turn so the agent can issue at most a single refund per conversation B) Keep issue_refund and add a system prompt rule saying the agent must never issue a refund without approval C) Keep issue_refund and add a PostToolUse hook that notifies the approver about each refund as soon as it is issued D) Remove issue_refund from the agent and give it a propose_refund tool that only writes a pending request to the approval queue

from claude_agent_sdk import ClaudeAgentOptions

options = ClaudeAgentOptions(
    mcp_servers={"billing": billing_server},
    allowed_tools=[
        "mcp__billing__lookup_order",
        "mcp__billing__propose_refund",
    ],
    disallowed_tools=["mcp__billing__issue_refund"],
)

Q274. A custom agent loop built on the Messages API triages support tickets. Its registry currently offers eight tools, including refund_order, close_account and reset_password, and a new tool is added most months. The triage job needs only read_ticket, search_kb and add_label. A security review asks for a configuration that stays safe as the registry grows. Which one? A) Build the request's tools array from an explicit list of read_ticket, search_kb and add_label, so nothing else is ever offered B) Offer every registry tool except refund_order, close_account and reset_password, which are filtered out by name C) Offer every registry tool and log each call to a restricted tool so the security team can review misuse weekly D) Offer every registry tool and state in the system prompt that only the three triage tools may be used

TRIAGE_TOOLS = ["read_ticket", "search_kb", "add_label"]

def tools_for_triage(registry):
    return [registry[name] for name in TRIAGE_TOOLS]

response = client.messages.create(
    model=MODEL,
    max_tokens=1024,
    tools=tools_for_triage(REGISTRY),
    messages=messages,
)

import os
from typing import Annotated

import anthropic
from anthropic import beta_tool
from pydantic import Field

from app.tickets import service  # your own ticket service

client = anthropic.Anthropic()  # reads ANTHROPIC_API_KEY
MODEL = os.environ.get("CLAUDE_MODEL", "claude-sonnet-5")


@beta_tool(name="add_label", description="Add one triage label to a ticket.")
def add_label(
    ticket_id: Annotated[str, Field(description="The ticket id, e.g. T-2291.")],
    label: Annotated[str, Field(description="One label from the triage list, e.g. billing.")],
) -> str:
    return service.add_label(ticket_id, label)


# read_ticket, search_kb, close_ticket, issue_refund are defined the same way
REGISTRY = {t.name: t for t in [read_ticket, search_kb, add_label, close_ticket, issue_refund]}
TRIAGE_TOOLS = ["read_ticket", "search_kb", "add_label"]


def triage(messages: list) -> object:
    runner = client.beta.messages.tool_runner(
        model=MODEL, max_tokens=1024,
        tools=[REGISTRY[name] for name in TRIAGE_TOOLS],  # only these exist for this agent
        messages=messages,
        max_iterations=8,
    )
    return runner.until_done()

{
  "name": "add_label",
  "description": "Add one triage label to a ticket.",
  "input_schema": {
    "type": "object",
    "properties": {
      "ticket_id": {"type": "string", "description": "The ticket id, e.g. T-2291.", "title": "Ticket Id"},
      "label": {"type": "string", "description": "One label from the triage list, e.g. billing.", "title": "Label"}
    },
    "required": ["ticket_id", "label"],
    "additionalProperties": false
  }
}

Drill · Prompting & Structured Output (Q275–Q296)

Q275. A hospital network summarizes discharge letters into a structured record. The record schema declares fasting_glucose as a required number, and the extraction runs with strict schema enforcement. Clinicians notice that letters for patients who never had a glucose test still come back with believable values such as 5.4. What is the root cause? A) The schema offers no legal way to express "not stated", so a required number field pushes the model to supply one; allowing null and telling it when to use null removes that pressure. B) Sampling temperature is too high for clinical data; setting it to zero would make the model stop producing values that are not in the letter. C) Strict schema enforcement is not actually active, because a working strict mode would reject any value that does not appear in the discharge letter. D) The discharge letters are too long, so the glucose section falls outside the attention of the model and a larger context window would fix it.

{
  "type": "object",
  "properties": {
    "fasting_glucose": {
      "type": ["number", "null"],
      "description": "mmol/L exactly as reported. Use null if the letter does not report a fasting glucose result."
    }
  },
  "required": ["fasting_glucose"],
  "additionalProperties": false
}

Q276. A ticket-routing service defines exactly one tool, classify_ticket, and leaves tool_choice at its default. About one ticket in forty comes back as a friendly paragraph instead of a tool call, which crashes the router. An engineer changes tool_choice to {"type": "tool", "name": "classify_ticket"}. Which statement about this change is exactly right? A) It makes every response a classify_ticket call instead of prose; with strict: true on the tool the arguments also match its input_schema, but the chosen category can still be wrong. B) It guarantees schema-valid arguments on its own, so the router can skip both strict mode and any validation of the category values it receives. C) It has the same effect as a system prompt line saying "always call classify_ticket", because tool_choice is only a strong hint that the model may still ignore. D) It guarantees the classification values are correct, because a forced tool call makes the model fill the schema from the ticket text rather than guess.

{
  "model": "claude-model-id",
  "max_tokens": 512,
  "tools": [{
    "name": "classify_ticket",
    "description": "Record the category and urgency of one support ticket.",
    "strict": true,
    "input_schema": {
      "type": "object",
      "properties": {
        "category": {"type": "string", "enum": ["billing", "login", "shipping", "other"]},
        "urgency": {"type": "string", "enum": ["low", "normal", "high"]}
      },
      "required": ["category", "urgency"],
      "additionalProperties": false
    }
  }],
  "tool_choice": {"type": "tool", "name": "classify_ticket"},
  "messages": [{"role": "user", "content": "Ticket: I was charged twice this month."}]
}

Q277. A freight forwarder extracts customs data from shipping manifests and retries any record that fails validation, feeding back the error each time. The operations lead wants to stop wasting calls on records that will never succeed. Which failures should be taken out of the retry loop and routed elsewhere? (Select TWO.) A) The tariff code field says "see customs annex", and the annex is a separate file that the pipeline never attaches to the request. B) The manifest template has no consignee phone number anywhere, and the schema still asks for one on every record. C) The response stopped with stop_reason max_tokens halfway through the list of containers, so the JSON never closed. D) The model wrote the gross weight in pounds into the field that expects kilograms, and the validator flagged the implausible value.

Q278. A recruiting platform parses résumés into employment histories. When a record fails validation, the pipeline calls the model again with the same résumé and the same instructions, and the second attempt succeeds only slightly more often than the first. Most failures are end dates written into the start_date field. What is missing from the retry request? A) A higher temperature on the second call, so that the retry explores a different interpretation of the résumé instead of repeating the first one. B) The failed extraction together with the validator message stating which positions have start and end dates reversed, so the model knows what to correct. C) A route to a larger model for every retry, since a model that failed once on a résumé is likely to fail again on the same text. D) Moving the formatting rules from the user message into the system prompt, where the model gives them more weight on the second attempt.

Q279. An insurer processes claims in two model steps: one assigns the claim to auto, home or travel, and a second extracts the fields that category needs. Travel claims for luggage stolen from a rental home are sometimes routed as home claims, so the second step extracts the wrong fields. A teammate suggests collapsing both steps into one large prompt. Which change most directly addresses the failure? A) Run the routing step three times and use the majority category, so an occasional misroute is outvoted before extraction. B) Send the extraction step to a larger model, so it can notice that the fields it was asked for do not match the claim. C) Keep both steps and add a check between them that tests the assigned category against evidence in the claim, rerouting or flagging it before extraction runs. D) Collapse both steps into one prompt as suggested, since a single call cannot pass a wrong category from one step to the next.

Q280. A property portal extracts floor area from listings. When the listing has a specification table the area comes back correctly, but when the agent writes it into prose such as "just shy of two thousand square feet over two levels", floor_area returns null. The instructions already define floor_area clearly. What should you try first? A) Make floor_area required and non-nullable, so the model can no longer return null when the area is only described in prose. B) Add a few examples where the area appears in prose, including approximate and spelled-out phrasings, each paired with the correct floor_area value. C) Raise max_tokens so the model has more room to reason through long prose descriptions before it fills in the record. D) Add a regular-expression preprocessor that converts any sentence mentioning square feet into a specification table before the call.

Q281. A staffing agency extracts weekly timesheets with tool use and checks each result against a Pydantic model. Every record validates and every field has the right type, yet payroll finds weeks where the seven daily hours do not add up to the weekly_total on the same record. Which statement is exactly right? A) Validation is silently failing, because a model that validates cleanly cannot contain records whose numbers disagree with each other. B) The Pydantic model only checks types, so no rule relates daily hours to weekly_total; add a validator for that sum and feed its error back into a retry. C) Adding ge=0 and le=168 bounds to weekly_total would reject these records, because disagreeing totals fall outside a plausible weekly range. D) Switching the extraction to strict: true would make the daily hours and weekly_total consistent, because strict mode checks the schema more thoroughly.

from pydantic import BaseModel, model_validator

class Timesheet(BaseModel):
    daily_hours: list[float]
    weekly_total: float

    @model_validator(mode="after")
    def totals_agree(self):
        total = round(sum(self.daily_hours), 2)
        if abs(total - self.weekly_total) > 0.05:
            raise ValueError(
                f"daily_hours sum to {total} but weekly_total is {self.weekly_total}; "
                "re-read the form and correct whichever value was misread"
            )
        return self

import os

import anthropic
from pydantic import BaseModel, ValidationError, model_validator

from app.review import review_queue  # your own module

client = anthropic.Anthropic()  # reads ANTHROPIC_API_KEY
MODEL = os.environ.get("CLAUDE_MODEL", "claude-sonnet-5")


class Timesheet(BaseModel):
    daily_hours: list[float]
    weekly_total: float

    @model_validator(mode="after")
    def totals_agree(self):
        total = round(sum(self.daily_hours), 2)
        if abs(total - self.weekly_total) > 0.05:
            raise ValueError(f"daily_hours sum to {total} but weekly_total is {self.weekly_total}; "
                             "re-read the form and correct whichever value was misread")
        return self


def extract(form_text: str, max_attempts: int = 2) -> Timesheet | None:
    messages = [{"role": "user", "content": f"Extract the timesheet.\n<form>{form_text}</form>"}]
    for _ in range(max_attempts):
        try:
            response = client.messages.parse(model=MODEL, max_tokens=1024,
                                             messages=messages, output_format=Timesheet)
            return response.parsed_output
        except ValidationError as err:  # the validator's message names both numbers
            messages = messages[:1] + [{"role": "user", "content": f"Your last answer failed: {err}"}]
    review_queue.put(form_text)  # the form itself may be inconsistent
    return None

Q282. A pharmacy system extracts dosing from prescriptions. Two problems remain: instructions such as "1 tab bid x 10" are read inconsistently, sometimes as ten days and sometimes as ten tablets, and quantity comes back empty on faxed prescriptions where it sits in a margin note. Which few-shot practices fit these problems? (Select TWO.) A) Add a handful of examples of the ambiguous dosing pattern, each with a short note explaining which reading was chosen and why. B) Replace the examples with a longer written definition of every field, since precise definitions remove ambiguity without extra prompt length. C) Include one example for every prescription template the pharmacy has ever received, so that no layout is left unrepresented in the prompt. D) Include examples drawn from different layouts, such as a faxed form with the quantity in a margin note, paired with the correct extraction.

Q283. A podcast network turns episode transcripts into show notes through a write_show_notes tool with tool_choice left on auto, and occasionally receives a written summary instead of a tool call. In the design review, four alternatives are proposed. Which one closes the gap with an actual guarantee? A) Move the sentence "always respond by calling write_show_notes" from the system prompt into the final user message, where it is read last. B) Prefill the assistant turn with an opening brace so the model has to continue a JSON object rather than start a summary. C) Keep the tool and set tool_choice to {"type": "tool", "name": "write_show_notes"}, so the API only allows a call to that tool. D) Drop the tool and ask for JSON in the prompt with a strict format block, since the model evidently prefers writing text for these transcripts.

Q284. A travel app extracts itineraries from booking confirmation emails through a record_flight tool. Validation often fails because departure_airport and arrival_airport are reversed relative to the "LHR → JFK" route line in the email. Today the retry just resends the email. Which follow-up turn gives the retry the best chance of success? A) A tool_result for the previous call with is_error set, stating that the route line reads LHR → JFK but the fields were recorded as JFK to LHR. B) No retry at all; application code swaps the two airport fields whenever the arrival time on the record is earlier than the departure time. C) A user turn that says only "That extraction was incorrect. Please try again more carefully." and resends the confirmation email. D) Five further identical calls with the same email, keeping the first result that passes validation, since the swap only happens on some attempts.

{
  "role": "user",
  "content": [{
    "type": "tool_result",
    "tool_use_id": "toolu_01A",
    "is_error": true,
    "content": "Validation failed: the email route line reads LHR → JFK, but departure_airport=JFK and arrival_airport=LHR. Re-read the route line and call record_flight again with the corrected airports. Other fields passed."
  }]
}

Q285. An expense system extracts reports and validates them. On one report the receipts sum to $212.40 but the claimed total reads $221.40. On another, the cost-center box on the form is blank and no cost center appears anywhere else. How should each be handled? A) Send both straight to human review without any retry, since any validation failure on financial data needs a person. B) Retry both, sending the discrepancy for the first and instructing the model to look harder for the cost center on the second. C) Retry the first with the specific discrepancy so the model rechecks receipts and total, flagging it if the source truly disagrees; send the second to review with cost center null. D) For the first, add a "miscellaneous" line of $9.00 so the report balances; for the second, fill in the employee's default cost center.

Q286. A biotech firm extracts fields from 120-page clinical study protocols in one request. Sponsor and phase from the opening pages come back reliably, but dosing schedules and stopping rules buried in later appendices are often missing, and the visit schedule sometimes contradicts the dosing table in the same output. The whole protocol fits comfortably in the context window. What is the best diagnosis and fix? A) Output varies between runs; run the full single-pass extraction three times and keep the most common value for every field. B) Attention is spread across too many fields and pages in one pass; extract per section in focused calls, then run a reconciliation pass that resolves conflicts. C) The protocol is near the window limit and later pages are being truncated; move to a model with a larger context window. D) The model does not realize dosing matters; add capitalized emphasis on dosing and stopping rules at the top of the existing single prompt.

Q287. An accessibility review bot comments on front-end pull requests. Teams mute it because it floods reviews with opinions about color palettes, button wording and spacing, burying real problems like missing alt text. The prompt was changed to "flag only significant accessibility issues" and nothing improved. Why did that change fail, and what would work? A) "Significant" has no boundary the model can test; list what to report, such as missing alt text or keyboard traps, and what to skip, such as palette taste. B) The bot should attach a confidence score to each comment and the pipeline should suppress anything the model rates below 0.8. C) The bot needs a hard cap of three comments per pull request, so the most important findings are forced to the top of the list. D) The model is too creative for review work; lowering temperature would make it produce fewer, more consistent comments per pull request.

Q288. Claude is writing a converter from a bank's legacy CSV export into a double-entry ledger JSON. Each revision fixes one problem and breaks another: refunds end up as debits, and the counterparty lands at the top level instead of inside the entry object. The mapping has been described in two paragraphs of prose. What should be supplied next? A) A request that Claude first restate the full mapping in its own words so any misunderstanding surfaces before the next revision. B) A JSON Schema for the ledger entry, with an instruction to validate every converted object against it before returning. C) An expanded specification listing every CSV column, its sign convention and its target path in the ledger object. D) Three real CSV rows, including a refund and a transfer, each paired with the exact ledger JSON object it should produce.

{
  "input_csv_row": "2024-02-03,-45.00,REFUND,ACME Outdoor",
  "expected_ledger_entry": {
    "date": "2024-02-03",
    "entry": {
      "type": "credit",
      "amount_cents": 4500,
      "counterparty": "ACME Outdoor",
      "memo": "refund"
    }
  }
}

Q289. A platform team has Claude generate reference entries for internal API endpoints, which a portal renders directly. Across months of use, about a fifth of entries omit the authentication requirement or the list of error responses, even though the system prompt describes both in detail. Every endpoint has an auth mode and at least one error response. Which change guarantees these fields are present? A) Record the documentation standard in the repository instructions file so it applies to every generation session without repeating it in each prompt. B) Add five complete example entries covering different endpoint styles, so the model sees authentication and error responses in every example. C) Define a record_endpoint_doc tool with auth_mode and error_responses as required fields, set strict: true, and force that tool for each entry. D) Run a linter over each generated entry and regenerate any entry that is missing the authentication or error response sections.

{
  "name": "record_endpoint_doc",
  "description": "Record reference documentation for one API endpoint.",
  "strict": true,
  "input_schema": {
    "type": "object",
    "properties": {
      "path": {"type": "string"},
      "auth_mode": {"type": "string", "enum": ["none", "api_key", "oauth", "mtls"]},
      "error_responses": {
        "type": "array",
        "items": {
          "type": "object",
          "properties": {"status": {"type": "integer"}, "meaning": {"type": "string"}},
          "required": ["status", "meaning"],
          "additionalProperties": false
        }
      }
    },
    "required": ["path", "auth_mode", "error_responses"],
    "additionalProperties": false
  }
}

Q290. An SRE team uses Claude to draft incident postmortems as Markdown pages for its wiki. Reviewing past drafts, customer impact is missing from roughly two in five. The prompt says "write a thorough postmortem", and leadership requires customer impact in every postmortem, even minor ones. What is the root cause of the gap? A) The output should become JSON with a required customer_impact field, because only schema enforcement can make a section appear reliably. B) "Thorough" never says which sections are mandatory, so the model decides case by case; name each required section and require it even when the finding is "none". C) The prompt lacks examples; adding two sample postmortems that include customer impact is the only way to establish the expected structure. D) The drafts need a larger model, because deciding what belongs in a thorough postmortem requires stronger judgment than the current one provides.

Q291. An online retailer's support pipeline classifies each ticket's product_area with a required enum: billing, login, shipping, returns. New kinds of tickets, such as gift card balances and screen-reader problems, are being filed as billing or login, and those queues quietly receive work they cannot handle. Which schema change keeps routing reliable and stops the forced matches? A) Replace the enum with a free-text string so the model can describe any product area in its own words. B) Make product_area nullable so the model returns null for any ticket that does not match one of the four areas. C) Add "other" to the enum plus an other_detail string that the model fills when no listed area fits the ticket. D) Expand the enum to thirty areas covering every ticket type the support team can think of today.

{
  "type": "object",
  "properties": {
    "product_area": {"type": "string", "enum": ["billing", "login", "shipping", "returns", "other"]},
    "other_detail": {
      "type": ["string", "null"],
      "description": "When product_area is other, describe the actual area in a few words; otherwise null."
    }
  },
  "required": ["product_area", "other_detail"],
  "additionalProperties": false
}

Q292. A logistics company added error-feedback retries to its bill-of-lading extraction and fixed most failures. Of the records that still fail, some ask for a seal number the document never contains, some repeat the same error of reading "freight collect" as a payment date, and some retries break fields that were already correct. Which plan matches each group? A) Accept null and queue the missing seal numbers for review; add targeted examples for the "freight collect" error; retry only the failing field, keeping validated fields fixed. B) Raise the retry limit to five for all three groups and write more detailed error messages, giving the model more chances to correct every failure. C) Stop retrying altogether and route every remaining failure to manual review, so nothing unresolved reaches the carrier systems. D) Make every field nullable across the schema; send the recurring error to a larger model; raise temperature on retries so they explore different outputs.

Q293. A food-labeling startup extracts nutrition panels with a strict tool schema, so every result parses and every field has the right type. Auditors still find panels where calories do not match the fat, carbohydrate and protein amounts, and sodium recorded in the potassium field. A developer proposes moving the schema from the tool into output_config.format. Which statement is exactly right? A) Both routes guarantee valid shape and types only; recompute calories from macros, cross-check mineral placement, and retry with the specific error. B) Adding more required fields, such as a calories_check field, would force the model to reconcile the macros and place each mineral correctly. C) Strict mode is being bypassed, because a working strict schema would reject a panel whose calories disagree with its macronutrients. D) Moving the schema to output_config.format would fix the errors, because structured outputs check values against the source while tool schemas only check types.

Q294. A commercial landlord extracts lease terms into a portfolio model. The schema requires annual_escalation_pct as a number. About a quarter of leases say only "rent escalates per the CPI schedule in Addendum C", and the addendum is not part of the file; for those, the model returns figures such as 3.0. Which schema change and instruction fix this? A) Add an escalation_basis text field next to the number so the model can record "CPI per Addendum C" when no percentage is given. B) Keep the field required and add an extraction_confidence number so the portfolio model can down-weight escalations the model is unsure about. C) Instruct the model to enter 0.0 when the escalation is defined elsewhere, and add a needs_addendum boolean set to true for those leases. D) Allow annual_escalation_pct to be null, and instruct the model to return null unless a percentage is explicitly stated in the lease text provided.

{
  "type": "object",
  "properties": {
    "annual_escalation_pct": {
      "type": ["number", "null"],
      "description": "Percentage stated in this lease text. Return null if the lease refers to another document for the rate."
    },
    "escalation_reference": {
      "type": ["string", "null"],
      "description": "Quote the reference, such as CPI schedule in Addendum C, when the rate is defined elsewhere; otherwise null."
    }
  },
  "required": ["annual_escalation_pct", "escalation_reference"],
  "additionalProperties": false
}

Q295. A bank's onboarding agent must call verify_identity_document before anything else, then choose among sanctions_screen, address_check and open_account, and it must never reply to the orchestrator in prose between steps. Sometimes it jumps straight to address_check, and sometimes it writes a status paragraph instead of calling a tool. Which tool_choice sequence enforces both requirements? A) On the first turn, force {"type": "tool", "name": "verify_identity_document"}; on later turns, use {"type": "any"} so some tool must be called. B) On the first turn, use {"type": "any"}; on later turns, force each specific tool by name in the order the compliance manual lists. C) Use {"type": "any"} on every turn, since that already requires a tool call and the model will naturally start with verification. D) Keep tool_choice on auto and add a hook that rejects any turn in which verify_identity_document was not the first tool called.

{
  "turn_1": {"tool_choice": {"type": "tool", "name": "verify_identity_document"}},
  "turn_2_onward": {"tool_choice": {"type": "any"}},
  "stop_condition": "orchestrator ends the loop after open_account returns or a check fails"
}

Q296. A shipping marketplace extracts parcel weight from seller listings into weight_kg. Sellers write "1.2 kg", "1,200 g", "2 lb 8 oz" and occasionally "two and a half pounds". A conversion script after extraction handles the metric forms but breaks on the imperial and spelled-out variants, and new variants keep appearing. What is the more robust fix? A) Instruct the model to report weight_kg in kilograms, with one short example for each known format, including a spelled-out imperial weight. B) Add a schema pattern that only accepts plain numbers in weight_kg, so any unconverted value fails validation and triggers a retry. C) Extend the conversion script with explicit patterns for every format seen so far, including the spelled-out imperial ones. D) Make weight_kg nullable so the model returns null for any listing written in a format that the script cannot convert reliably.

Drill · Tools & MCP (Q297–Q317)

Q297. An agent is moving a service's config file off a retired database host. In deploy/values.yaml the string db-old.corp.internal appears 31 times, and every one of them must become db-new.corp.internal. The agent calls Edit with that old string and the new one, and Edit refuses because the old string is not unique in the file. What is the right next call? A) Call Edit 31 times, each with enough neighbouring lines in old_string that only one occurrence matches per call B) Glob for '**/db-old' to find the files involved, then Edit each file it returns C) Repeat the same Edit call with replace_all set to true, since every occurrence in this file is meant to change D) Read the whole file, then Write it back with the new host typed into each of the 31 places

{
  "tool": "Edit",
  "input": {
    "file_path": "deploy/values.yaml",
    "old_string": "db-old.corp.internal",
    "new_string": "db-new.corp.internal",
    "replace_all": true
  }
}

Q298. A fare-comparison orchestrator sends a subagent to collect prices for 40 flight routes. At route 26 the airline portal starts rejecting requests because the subagent's session token has expired. The orchestrator decides whether to refresh credentials, reassign work, or report a partial comparison. Which TWO things should the subagent's error report contain? (Select TWO.) A) A permission failure category (expired session) and the request that was rejected B) A decision to abandon the whole comparison, since the subagent is the component closest to the failure C) The prices already collected for routes 1–25, plus the list of routes still uncovered D) A success status carrying the 25 routes, so the comparison is not marked as failed

Q299. In a 2,000-line legacy test file, Claude Code needs to change the timeout inside one specific retry block. The same five-line block is pasted in several tests, so Edit returns a match-not-unique error. Only the block inside test_refund_partial should change. Which recovery changes exactly that block without restructuring the file? A) Repeat the Edit with replace_all set to true so the uniqueness refusal no longer triggers at all B) Move each test into its own file so every copy of the block becomes unique per file C) Read the full file, change that one block in memory, and Write it back D) Use Bash with sed to rewrite the timeout on a line number read from an earlier grep

Q300. A telehealth intake assistant must call record_consent before any clinical conversation starts. Afterwards it should answer questions freely, calling other tools only when useful. In testing, the first reply is sometimes a friendly greeting with no tool call. Which request setting guarantees the consent call on the opening request without constraining later turns? A) Send only record_consent in the opening request's tools list and leave tool_choice on auto B) Keep tool_choice on auto and add a system prompt line saying consent must always be recorded first C) Use tool_choice type any on every request so some tool is always called before any text is produced D) Opening request: tool_choice naming record_consent; every later request: auto

{
  "opening_request": {
    "tool_choice": {
      "type": "tool",
      "name": "record_consent"
    }
  },
  "later_requests": {
    "tool_choice": {
      "type": "auto"
    }
  }
}

Q301. A pharmacy assistant checks refill status with lookup_prescription. For one caller it gets an empty list, reformats the prescription number and tries four more times, then pages a pharmacist. The pharmacist finds the number was never issued. The same empty list is also what the tool returns when the pharmacy database is unreachable. What is the root cause of the wasted retries and the page? A) The model's sampling temperature is too high, which makes it keep trying new number formats B) The escalation policy pages a pharmacist too early and should allow several more lookup attempts first C) The retry limit is set too high; cutting it to two attempts would have avoided most of the wasted work D) Its result gives a valid no-match and a failed lookup the same shape, so they look identical

Q302. A localization agent pushes 400 UI strings through an MCP translation server. Around string 120 the upstream vendor starts answering HTTP 429 with a Retry-After of 45 seconds, and the server passes back only 'Translation failed'. The agent concludes the service is broken and stops. Which server-side change lets the agent finish the batch on its own? A) Hold each call open inside the server, retrying quietly until the vendor accepts, however long that takes B) Mark the failure as a permission error that is not retryable, so the agent stops sending requests to the vendor C) Return isError true, category transient, retryable true, and a wait hint from Retry-After D) Return the source string untranslated for any string that hits the limit, so the batch always completes