The 6 Exam Scenarios

A. Escalate immediately because of the negative sentiment. B. Resolve the order-status request; sentiment alone is not an escalation trigger. C. Escalate because the model’s confidence is below 70%. D. Cap the conversation at 3 turns then escalate.

Answer: B. Sentiment is not a trigger (#5); the request is within capability, so resolve it. Confidence-based escalation is #4, and an arbitrary cap is not an escalation rule.

A. A sentence in the agent’s system prompt. B. A PreToolUse hook on issue_refund that inspects the amount and exits 2 to block, routing to a human. C. Asking the model to double-check itself. D. A note in CLAUDE.md.

Answer: B. Critical rules need deterministic hooks (#3). Prompt/CLAUDE.md text (A, D) and self-check (C) are probabilistic.

A. Longer backoff. B. An idempotency key on issue_refund so retries do not duplicate the effect. C. Never retrying anything. D. A bigger model.

Answer: B. Idempotency keys make write retries safe. Backoff (A) does not prevent duplication, never retrying (C) is unnecessary, and model size (D) is irrelevant.

A. Add it to project CLAUDE.md. B. A checked-in PreToolUse hook that runs tests on commit and exits 2 on failure. C. Ask each developer to remember. D. A slash command that runs tests.

Answer: B. A checked-in hook enforces it deterministically for everyone. CLAUDE.md (A) and memory (C) are probabilistic; a slash command (D) is opt-in.

A. The team’s build and test commands. B. A developer’s personal scratch notes and local machine paths. C. The repo’s architecture conventions. D. A mandatory security rule.

Answer: B. Personal, non-shared notes go in the git-ignored local file. Team commands/conventions (A, C) go in project CLAUDE.md; mandatory rules (D) go in managed policy.

A. Direct execution to move fast. B. Plan mode: read-only exploration then a reviewable plan before edits. C. Delete failing tests. D. Increase max_tokens.

Answer: B. Large, unfamiliar, multi-file work is the canonical plan-mode case. Direct execution (A) skips review, deleting tests (C) is destructive, and max_tokens (D) is irrelevant.

A. The subagent needs a bigger window. B. Subagent contexts are isolated and do not inherit findings; the coordinator must pass context explicitly. C. The subagent used the wrong model. D. Thinking was disabled.

Answer: B. Isolated contexts never auto-inherit. Window size (A), model (C) and thinking (D) do not supply context that was never passed.

A. Present the three results as the complete answer. B. Record structured errors (timeout, retryable), retry the retryables, then proceed with a quorum noting the gap or escalate. C. Discard everything and restart. D. Return ‘research failed’.

Answer: B. Structured errors plus an explicit decision. Silent drop (A) is #7, restart (C) wastes good work, and a generic message (D) is #6.

A. Move to Haiku 4.5 for a bigger window. B. Have subagents return only distilled, structured results; their isolated contexts hold the raw material. C. Turn off thinking. D. Remove error handling.

Answer: B. Subagent isolation plus distilled returns keep the main window small. Haiku 4.5 (A) has a smaller (200k) window, and C/D are unrelated or harmful.

A. Code execution. B. Web search server-side tool. C. Memory tool. D. Computer use.

Answer: B. Web search grounds with citations. Code execution (A) runs code, memory (C) persists state, computer use (D) drives a desktop.

A. Three separate custom tools. B. An MCP server exposing the capability, connected by each host and via the Messages API MCP connector. C. A slash command. D. A CLAUDE.md note.

Answer: B. MCP provides one reusable cross-client integration. Three tools (A) duplicate work; a slash command (C) and CLAUDE.md (D) are Claude Code artefacts, not integrations.

A. Yes, for consistency. B. No — call the API/CLI directly; wrapping it as a model tool needlessly adds latency, cost and non-determinism. C. Yes, expose it as an MCP resource. D. Only if it is on Fable 5.1.

Answer: B. Deterministic steps your code owns should be called directly. Wrapping (A, C) is over-engineering; the model version (D) is irrelevant.

A. claude -p "…" --output-format json --allowedTools "Read,Grep,Bash(git diff:*)", parse JSON, gate on exit code. B. Interactive claude, copy results by hand. C. --permission-mode bypassPermissions with all tools. D. Plain claude -p and grep the prose.

Answer: A. Headless JSON output, minimal allowlist, exit-code gating. Interactive (B) does not automate, bypassing permissions (C) is unsafe, and grepping prose (D) is unreliable.

A. Real-time Messages API at high concurrency. B. The Message Batches API — 50% discount, results within 24h, latency-tolerant. C. One giant request with all 500 diffs. D. Force tool_choice on Fable 5.1.

Answer: B. Batch API fits latency-tolerant bulk work at half price. High concurrency (A) risks limits and costs more, one request (C) will not fit, and forcing tool_choice on Fable 5.1 (D) 400s.

A. Truncate to 500 lines. B. Multi-pass: partition by module, review each in its own pass/subagent, then aggregate and rank findings. C. One giant prompt with the whole diff. D. Skip review.

Answer: B. Partition-review-aggregate preserves quality. Truncation (A) misses code, one giant prompt (C) degrades quality, and skipping (D) is unacceptable.

A. Retry with backoff. B. Use tool_choice: 'auto' with an instruction to call the tool, strict: true schemas, or structured outputs. C. Use tool_choice: 'any'. D. Lower max_tokens.

Answer: B. Fable 5.1 forbids forced tool choice; use auto+instruction, strict, or structured outputs. Backoff (A) does not fix a 400, any (C) is also blocked, and max_tokens (D) is unrelated.

A. Increase the sample size. B. Report per-document-type accuracy and gate on the worst type. C. Raise temperature for contracts. D. Average more runs.

Answer: B. Aggregate metrics hide a failing type (#10); per-type metrics expose it. Larger samples (A) still aggregate, temperature (C) does not fix accuracy, and averaging (D) hides the problem more.

A. More blind retries. B. Feed the specific validation error back to the model and require valid JSON matching the schema; parse defensively. C. Use eval() to parse. D. Grade the output in the same session that produced it.

Answer: B. Specific error feedback drives self-correction. Blind retries (A) rarely help, eval() (C) is unsafe, and same-session grading (D) is #9.

A. Remove the timeout so slow calls eventually return. B. Set a boundary timeout; on timeout return a structured {category:'timeout', retryable:true} error so the agent can retry, proceed noting the gap, or escalate. C. Catch the timeout and return an empty success. D. Use a larger model so the tool responds faster.

Answer: B. Boundary timeouts plus a structured error keep the agent responsive and honest. No timeout (A) lets a hung tool stall the agent, empty success (C) is #7, and model size (D) does not change a downstream tool’s latency.

A. A confidence threshold on the model. B. A deterministic PreToolUse hook on the transfer tool that inspects the amount and exits 2 above the threshold, ignoring any injected ‘pre-approval’. C. Escalate only if the customer seems anxious. D. A system-prompt instruction to ask before large transfers.

Answer: B. Irreversible high-value actions need a deterministic gate keyed on the amount, immune to injection. Confidence (A) is #4, sentiment (C) is #5, and a prompt rule (D) is #3.

A. PostToolUse to block the commit; PreToolUse to lint. B. PreToolUse on the commit (run tests, exit 2 on failure) and PostToolUse on Edit to lint. C. UserPromptSubmit for both. D. SessionStart to lint and Stop to run tests.

Answer: B. Blocking happens before the action (PreToolUse, exit 2); linting reacts after the edit (PostToolUse). A swaps the events, C uses a prompt event, and D fires at the wrong times.

A. Haiku 4.5 (user is most personal). B. Sonnet 5 — managed policy cannot be overridden. C. Opus 5 (project is closest to code). D. Whichever file changed last.

Answer: B. Managed policy overrides local, project and user. A and C invert precedence; D is not how resolution works.

A. Compaction, because it summarises everything. B. Context editing to clear the stale tool results while preserving the narrative; compaction would be wrong here because it summarises the dialogue rather than targeting bulky tool outputs. C. Switch to Haiku 4.5 for its larger window. D. Increase max_tokens.

Answer: B. Clearing bulky stale tool results is context editing, which keeps the dialogue. Compaction (A) targets the narrative, Haiku 4.5 (C) has a smaller 200k window, and max_tokens (D) is unrelated.

A. Workers need bigger windows. B. Context is isolated at every tier; each level must pass the relevant context explicitly to the next — inheritance never happens at any depth. C. The root should use Opus 5. D. Thinking is disabled at the worker tier.

Answer: B. Isolation applies at every tier, so context must be threaded explicitly all the way down. Window size (A) and model (C) do not supply un-passed context; thinking (D) is unrelated.

A. Both as Tools. B. The policy as a Resource (application-controlled data) and search_orders as a Tool (model-controlled action). C. Both as Prompts. D. The policy as a Tool and search_orders as a Resource.

Answer: B. App-supplied reference data is a Resource; a model-invoked action is a Tool. Making the policy a Tool (A, D) adds needless model decisions; Prompts (C) are user-invoked templates.

A. Full OAuth scopes so it never lacks a capability. B. Scope its OAuth 2.1 grant to read-only order access (least privilege applied to auth). C. Embed an admin API key in the prompt. D. Use stdio so no auth is needed.

Answer: B. Least privilege applies to MCP auth. Full scopes (A) widen blast radius, prompt-embedded keys (C) are insecure, and stdio (D) is a local transport, not an option for a shared remote server.

A. --output-format text, then grep for ‘HIGH’. B. --output-format json, parse findings, and fail on a non-zero exit code and any high-severity finding. C. --output-format stream-json and ignore the exit code. D. Interactive mode with a human reading output.

Answer: B. A single JSON result plus exit-code-and-findings gating is correct for a short task. Grepping prose (A) is unreliable, ignoring the exit code (C) misses failures, and interactive mode (D) does not automate.

A. Cache the stable system + schema prefix so reads cost ~0.1x. B. Use the Message Batches API for the bulk run (50% off, ≤24h). C. Call the real-time API at maximum concurrency. D. Use Fable 5.1 at $10/$50 for every document. E. Randomise the prompt each call.

Answer: A and B. Caching the fixed prefix plus the Batch API cut cost sharply for latency-tolerant bulk work. Real-time concurrency (C) is full price and limit-prone, the priciest model (D) raises cost, and randomising (E) destroys the cacheable prefix.

A. One loose object with everything optional. B. A discriminated union (oneOf with a const doc_type discriminator) so each branch enforces its own required fields. C. Three unrelated endpoints with no shared contract. D. A single string field holding raw JSON text.

Answer: B. A discriminated union enforces per-type requirements in one schema. A loosens everything, C loses a shared contract, and D abandons schema guarantees.

A. Write the partial object; it is mostly complete. B. Treat max_tokens as truncation: raise the limit or chunk the document (or bound arrays with maxItems), then retry — never persist the partial output. C. Return an empty object so the pipeline continues. D. Ask the model in the same session whether it finished.

Answer: B. max_tokens is truncation, not completion. Persisting the partial (A) corrupts data, empty (C) is #7, and same-session self-check (D) does not address truncation.