Skilljar 官方 Quiz 题库

You're building a multi-turn chatbot. The user sends a second message. How do you preserve context with the Anthropic API?

1. Re-send only the new user message; the API remembers the previous turn
2. Use a session_id query parameter to look up history server-side
3. Send the full prior message list (user + assistant turns) plus the new user message
4. Put the prior conversation inside the system prompt

Your `max_tokens=200` request returns a response with `stop_reason: "max_tokens"`. What happened?

1. Your input prompt exceeded 200 tokens and was truncated
2. The model reached the output budget before finishing its answer
3. Anthropic billing capped your account at 200 tokens for the day
4. The conversation context window is now full

You need to summarize 50,000 customer reviews per day, each ~500 words. Cost matters; quality bar is moderate. Which model is the right default?

1. Always Opus, because it has the strongest reasoning
2. Haiku, because it offers the best cost-latency profile for high-volume, well-bounded tasks
3. Whichever model has the largest context window
4. Sonnet, because it's the only model that supports streaming

Your team is reviewing a pull request that adds Claude API integration. Where should the API key live?

1. Hard-coded as a constant in the source file with a comment "rotate yearly"
2. Committed in `.env` so deployment is reproducible across machines
3. In an environment variable / secret manager, never committed to the repo
4. Embedded in the frontend bundle so the backend stays stateless

You set `stop_sequences=[""]`. The model generates text containing that string. What happens?

1. The API returns an error because stop sequences must not collide with content
2. Generation halts at the first match; `stop_reason` becomes `stop_sequence` and the matched string is NOT included in output
3. The matched string is included verbatim and generation continues
4. Stop sequences are silently ignored unless temperature is below 0.5

Why count tokens before sending a request to Claude?

1. It's optional — the API will reject anything that doesn't fit and retry shrinks automatically
2. Only required for streaming; non-streaming requests are unbounded
3. To stay within the model's input window AND to estimate cost before committing to the call
4. To choose the right `temperature` value

A prompt tweak improves Q1–Q3 of your eval set but silently regresses Q5. What is this telling you?

1. Q5 is noise; ignore single-question regressions
2. The model has a bug; switch to a different Claude version
3. Prompt changes have tradeoffs across distributions; only a full eval set surfaces regressions
4. You should revert Q1–Q3 improvements until Q5 stabilizes

Which grading style fits "the output JSON must validate against this schema"?

1. Model-based grading: ask Claude to judge schema validity
2. Code-based grading: parse JSON and run a schema validator
3. Manual review by the prompt author
4. Either — they produce equivalent results

When should you refresh your eval dataset?

1. Never — a stable eval set is the whole point
2. After every prompt change, to keep it aligned
3. When production data drifts or when new failure modes emerge from real users
4. Once a year, regardless of usage

You're building your first eval set. Which inclusion matters most?

1. The 10 prompts your team uses most often
2. A mix of common cases, edge cases, adversarial inputs, and real production samples
3. Synthetic inputs generated by Claude itself, for scale
4. Only inputs that the current prompt already handles correctly

Why log prompts and responses in production?

1. It's required by Anthropic for billing audits
2. To rebuild eval datasets with the real distribution of user inputs and to debug regressions
3. To train a custom Claude variant on your domain
4. To detect malicious users automatically

"Be brief" vs "Limit to 200 words, exactly 3 bullets, no markdown headers" — which works better and why?

1. Both are equivalent; Claude infers intent
2. "Be brief" is better; over-specifying confuses the model
3. The specific version, because measurable constraints replace ambiguous adjectives
4. Depends on temperature; higher temperature handles vague instructions

Few-shot works for your task but the examples eat too much budget. What's the next move?

1. Drop few-shot entirely
2. Switch to a higher temperature to make examples optional
3. Move stable examples to a cached prefix (prompt caching) so they amortize across requests
4. Concatenate all examples into a single short paragraph

Your prompt feeds Claude a 50-page document followed by 3 lines of instructions. Outputs sometimes drift off-task. What's the fix?

1. Always place instructions BEFORE the document, never after
2. Repeat the key instructions AFTER the document, near the end of the prompt
3. Shrink the document to fit instructions at the top
4. Increase temperature to escape document anchoring

When is asking Claude to "think step by step" the WRONG choice?

1. Multi-step math problems
2. Code review tasks
3. High-volume, latency-sensitive classification with simple inputs
4. Complex policy reasoning

"Act as a senior security auditor reviewing this code" — what's the actual mechanism behind role prompts?

1. Claude loads a different model fine-tuned for that role
2. It enables hidden tools tied to the role
3. It biases Claude's vocabulary, depth, and priorities toward that domain's writing patterns
4. It's purely cosmetic and has no measurable effect

"Clear and direct" 与 "specific" 在 prompt 工程中的区别是什么？

1. 它们是同一个原则的不同说法
2. Clear/direct 强调任务表述无歧义；specific 强调补充目标、受众、约束、格式与成功标准
3. Clear/direct 仅用于 system prompt；specific 仅用于 user prompt
4. Clear 是英文专属；specific 是中文专属

Claude returns a `tool_use` block with `id="toolu_xyz"`. Your code runs the tool. What do you send back?

1. A new user message saying "the tool returned X"
2. A system message updating Claude's instructions
3. A `tool_result` content block with `tool_use_id="toolu_xyz"` and the result payload, inside a user message
4. Nothing — Claude assumes success unless an error is raised

Your tool description reads: "gets data". Claude often picks the wrong tool. Most likely cause?

1. The model is too small — switch to Opus
2. The tool description is too vague to disambiguate from siblings
3. Tool descriptions don't influence selection — it's all from parameter names
4. You forgot to set `tool_choice="auto"`

Your refund agent has a `process_refund` tool that issues real money. Right gating pattern?

1. Auto-execute everything Claude calls — that's the point of agents
2. Hide the tool entirely; never expose write actions to Claude
3. Have Claude propose the refund; require human confirmation before the side effect runs
4. Run the tool but log it for retroactive audit

Your tool returns `{"error": "rate_limited"}`. What should Claude do next, in a well-designed loop?

1. Treat error as success and continue
2. Read the error, decide between retry, alternative path, or escalate to the user
3. Always retry immediately with the same args
4. Crash the entire conversation

When are parallel tool calls a good fit?

1. Always — it speeds up everything
2. Only with the code execution tool
3. When operations are independent, e.g. fetching weather AND fetching news for the same response
4. Never — Claude can't handle parallel results

A tool's `input_schema` requires fields `customer_id: string` and `since: ISO date`. Claude calls with `customer_id: 12345` (integer). What's the right fix?

1. Loosen the schema to accept any type
2. Switch to a smarter model
3. Tighten the description: clarify "customer_id is a string like 'cust_001', NOT a numeric ID"
4. Validate types only after Claude has executed the tool

Extended thinking — what does it actually do?

1. Replaces RAG by giving Claude internal knowledge access
2. Always produces faster responses than non-thinking mode
3. Allocates a visible reasoning budget before the final answer, helping on complex multi-step problems
4. Only available for code generation tasks

Prompt caching — what gets cached and how?

1. The final answer, keyed by the input hash
2. A stable prefix marked by a `cache_control` breakpoint; subsequent requests reuse it at reduced cost
3. Caching is automatic for any prompt over 1k tokens
4. Tool definitions, but not message content

You want Claude's answer to cite the source paragraph from your document. What's required?

1. Just write "cite your sources" in the prompt
2. Use higher temperature to force citation behavior
3. Use the Citations feature: pass the document with `citations: { enabled: true }` so Claude can attach citation blocks to spans of its output
4. Manually post-process Claude's text and match keywords to the document

When is the Files API a better fit than just inlining the file content into a message?

1. Whenever the file is over 100 KB
2. When the same large file will be referenced across many requests, so uploading once is cheaper than re-sending
3. When you need to encrypt the file end-to-end
4. When you want streaming output

Message Batches — main reason to use it?

1. Faster responses than streaming
2. Up to 50% cost reduction for async, non-time-sensitive workloads
3. Required for any workload above a certain RPS
4. Replaces streaming for chat applications

PDF + image input — which statement is true?

1. PDFs must be converted to plain text before sending
2. Images are free and don't count toward input tokens
3. Both are accepted natively as content blocks and consume input tokens proportional to size and pages
4. Vision capability requires a separate Anthropic Vision API endpoint

API tool use vs MCP — what's the primary difference?

1. MCP runs faster than API tool use
2. API tool use is deprecated and being replaced by MCP
3. MCP standardizes a server protocol so the same tools/resources/prompts can be reused across many clients without rewriting
4. They are identical; MCP is just marketing

Among MCP primitives — tools, resources, prompts — which is intended for read-only context fetch?

1. Tools
2. Resources
3. Prompts
4. Sampling

Your MCP server exposes internal company HR data. What MUST your design account for?

1. Trust the protocol to enforce access automatically
2. Default-allow then audit afterwards
3. Per-user authentication, scoped resource visibility, audit logs of every call
4. Disable HTTPS to simplify debugging

What role do MCP `prompts` play?

1. System prompts that the client must use verbatim
2. A way to override the client's tool definitions
3. Server-defined, reusable prompt templates the client can surface to users (e.g., "summarize this PR")
4. Client-side only — they don't cross the protocol

When is adding MCP overkill?

1. When the model is Sonnet
2. Whenever cost matters
3. When you have a single, tightly-coupled tool that no other client will ever consume — direct API tool use is simpler
4. Never — MCP is always the right choice

An attacker plants a malicious instruction inside an MCP `resource` returned to Claude. Best defense?

1. HTTPS alone — transport encryption stops injection
2. Trust Claude to ignore obvious instructions
3. Server-side validation of resource content + client-side instruction isolation (treat resource bodies as data, not commands)
4. Disable resources entirely; only allow tools

When does the workflow pattern beat the agent pattern?

1. When the task involves any tool calls
2. When the steps are known, predictable, and you need stable cost / latency / testability
3. When you're using a smaller model like Haiku
4. When the user prefers conversational responses

In an agent loop, why is "verify" a distinct step?

1. It's a billing milestone
2. To compress conversation history
3. To confirm the action achieved the goal and surface tool failures or unintended effects before continuing
4. To switch models mid-loop

What are sensible stop conditions for an agent loop?

1. Only when the goal is fully achieved — don't bound iterations
2. Only when an error occurs
3. Goal met OR max iterations reached OR explicit user halt OR cost / time budget exceeded
4. Stop after exactly 5 turns regardless

A pipeline stage is a fixed sequence of three deterministic API calls. How should you implement it?

1. Wrap it in an agent loop for "future flexibility"
2. Use extended thinking
3. Plain code orchestrating three Claude calls — workflow, no agent loop needed
4. Hardcode the calls inside the system prompt

For high-stakes verification, why use an INDEPENDENT Claude call rather than asking the same agent to self-check?

1. Independent calls are always faster
2. It's required by the API for tool use
3. Self-check shares context bias; an independent call sees only the artifact and gives a less anchored judgment
4. Self-checks are forbidden by Anthropic policy

How do you keep agent costs predictable in production?

1. Costs are inherently unpredictable; budget generously
2. Cap latency only — cost will follow
3. Cap iterations + log per-turn token usage + alert / escalate on cost spikes
4. Run only with Haiku regardless of task

A customer-support bot must answer questions from internal policy docs and never fabricate. Best architecture?

1. Pure agent loop with web search
2. Inline the entire policy into the system prompt every request
3. RAG over policy docs + Citations + an eval set that tests "I don't know" / escalation paths
4. Fine-tune a custom Claude on the policy docs

Compliance team needs to audit every Claude response against source documents. Best feature combo?

1. High temperature + manual review
2. Citations + structured output (JSON for audit fields) + full request/response logging
3. Just bigger model and trust the answers
4. Streaming, so reviewers see drafts in real time

You're generating SQL from natural-language questions. Which temperature setting is sane?

1. 1.0+, to encourage creative joins
2. 0.0–0.2, for deterministic, syntactically correct SQL
3. It doesn't matter for code generation
4. 0.7, the default that fits everything

Daily report job analyses the same 100k-token policy book against today's incidents. Cost optimization?

1. Switch all calls to Haiku regardless of quality needs
2. Truncate the policy book to 10k tokens
3. Use prompt caching with the policy book as a stable prefix — every call after the first hits the cached copy
4. Run streaming so partial results lower wall-clock time

An agent has a `process_refund` tool. The biggest production risk?

1. The tool description being too long
2. Using JSON instead of XML for the schema
3. Auto-executing refunds without a human-in-the-loop confirmation gate
4. Calling it from Sonnet instead of Opus

How do you detect prompt regressions across versions?

1. Wait for user complaints
2. Manual spot-check the new prompt on a few prompts you remember
3. Maintain a versioned eval set with consistent grading; run it on every prompt version and compare scores
4. Trust streaming output to surface anomalies

You need Claude to actually run Python code, not just generate it. Best mechanism?

1. Add "execute the code" to the system prompt
2. Spin up a separate agent that copies code from Claude's text output
3. Use the code execution tool — sandboxed runtime that returns real results to Claude
4. Switch to a bigger model

Your team exposes the company CRM via an MCP server. What governs which records each user can see?

1. The MCP protocol enforces row-level security automatically
2. The system prompt on the client side
3. Server-side authentication + per-user scope on resources/tools, enforced before responding to any MCP call
4. Trust by default; let Claude decide what to expose

What does a `CLAUDE.md` file in a project root actually do?

1. It's a README replacement that GitHub renders specially
2. It must be present for `claude` to launch in that directory
3. It's auto-loaded as persistent project memory — instructions, conventions, and file pointers stay visible to Claude across sessions
4. It overrides Claude's safety system prompt

You want Claude to look at `src/auth/login.ts`. Best way to add it as context?

1. Paste the entire file content into the prompt
2. Reference it with `@src/auth/login.ts` — Claude reads on demand and avoids bloating the context window
3. Copy the entire repo and let Claude scan everything
4. Disable context limits and dump the whole project

Your team keeps re-asking Claude to "review the diff for accessibility issues using our checklist." How do you make this a one-shot command?

1. Paste the checklist into every prompt
2. Edit the global system prompt
3. Create `.claude/commands/a11y-review.md` with the workflow + checklist; invoke via `/a11y-review`, supports `$ARGUMENTS` for parameters
4. Hard-code it in a shell alias

Adding a Linear MCP server so Claude Code can read tickets — where does the config go?

1. It's hard-coded inside the `claude` binary; you can't add servers
2. `.mcp.json` at the project root declares per-project servers (command, args, env); commit it for the team
3. Only `~/.claude/settings.json` works — MCP is global-only
4. Pass `--mcp` flag with the server URL on every launch

A `PreToolUse` hook in `settings.json` — what can it do?

1. Only log the tool call to stdout — purely cosmetic
2. Fire after the tool finishes, to record results
3. Run before the tool call; exit code or JSON output can deny, allow, or modify the call
4. Replace the entire Claude system prompt for one session

Running Claude Code in a GitHub Actions workflow to triage PRs — what's the right shape?

1. Run the interactive CLI; pipe stdin from the PR body
2. Skip Claude in CI; only run it locally
3. Use the Claude Agent SDK or GitHub Action: stateless invocation with the prompt + repo checkout, `ANTHROPIC_API_KEY` from secrets, output posted as PR comment / patch
4. Hard-code the API key into the workflow YAML

In MCP, who initiates the connection — host, client, or server?

1. The server connects out to the host on startup
2. The host (e.g., Claude Code) spawns or connects a client, which initiates the handshake with the server
3. Server and client connect peer-to-peer with no host involved
4. The order doesn't matter — MCP is symmetric

You're building an MCP server for a knowledge base. Users should be able to (a) search articles, (b) read full article text, (c) trigger a "draft a reply" workflow. Which primitives map to which?

1. All three should be tools — keep it simple
2. Search = tool (action with side-effect of result), full text = resource (read-only context), draft-reply workflow = prompt (named template)
3. All three should be resources — they all read data
4. All three should be prompts — they're triggered by user intent

MCP Inspector — primary purpose?

1. It's a deployment service that hosts your MCP server in production
2. A local debugging UI that connects to your server, lists its tools/resources/prompts, and lets you call them manually to verify behavior
3. An IDE plugin that auto-generates MCP server code
4. A logging aggregator for production MCP traffic

A tool's `input_schema` lists `{"query": "string", "limit": "integer"}` with no descriptions. What's wrong?

1. Nothing — types are sufficient for the LLM to call correctly
2. The schema must use `kind` instead of `type`
3. Without per-field descriptions, the LLM has no signal about valid ranges, defaults, or semantic meaning — quality degrades
4. Schemas don't matter; only the tool description does

In Claude Code, you want a Linear MCP server visible only to your team in this repo. Which scope and config?

1. User scope (`~/.claude.json`) — easiest to maintain
2. Project scope: commit `.mcp.json` in the repo root so every collaborator picks it up
3. Local scope: `claude mcp add --scope local`, then commit your machine's settings file
4. Hard-code it in CLAUDE.md so team members read about it

You're writing your own MCP client (not using Claude Code). What does the client need to do besides JSON-RPC plumbing?

1. Nothing — JSON-RPC handles everything
2. Negotiate capabilities at startup, surface tools/resources/prompts to the LLM, route LLM tool calls to the right server, and pass tool_result back into the conversation
3. Implement its own LLM internally; servers refuse plain HTTP clients
4. Encrypt every payload before sending

Sampling in MCP — what does it actually do?

1. The server sends statistics about LLM token usage to the host
2. The server requests the host's LLM to generate a completion on its behalf — capability negotiation must enable it, and the host approves before forwarding
3. It's a load-balancing technique for multiple MCP servers
4. It samples a subset of resources to avoid overloading the model

Roots — what are they for?

1. The set of administrators allowed to install MCP servers
2. Cryptographic root certificates for secure transport
3. Filesystem boundaries the host advertises to servers — "you may operate within these directories" — letting servers scope their work appropriately
4. The top-level routes in an HTTP-based MCP server

Picking transport: STDIO vs Streamable HTTP. Which fits "internal Linear MCP server hosted on a private VM, used by 50 employees with team-shared auth"?

1. STDIO — keeps things simple
2. Streamable HTTP — multi-user remote access requires a network transport with auth and session handling
3. Either works equivalently
4. Neither — MCP can't serve multiple users

Your MCP tool runs a 90-second build. Without progress notifications, what happens on the host side?

1. The host shows "running..." indefinitely with no feedback — users assume it's stuck and may abort
2. Streaming HTTP automatically emits keep-alive — no action needed
3. JSON-RPC times out at 30s and fails the call
4. The tool can't run longer than 30s in MCP at all

Streamable HTTP — stateless vs stateful session. When should you go stateful?

1. Always — stateful is more reliable
2. Never — stateless scales better
3. When the server holds per-session resources (in-flight cursor, partial uploads, conversation memory) that can't be reconstructed from a single request
4. Whenever you have more than 10 concurrent users

Productionizing an HTTP MCP server. Which is the most important security control?

1. Choose a unique port number to obscure the server
2. Authenticate every request, scope tools/resources per identity, audit log every call, and treat all external content as untrusted (injection defense)
3. Disable HTTPS for performance — internal traffic is safe
4. Run with default settings; MCP includes built-in security

Three tasks land on your desk. Which one is the strongest fit for Cowork (vs. Chat or Claude Code)?

1. Quickly summarize a single article you pasted into the chat
2. Refactor a Python module's tests in a git repository
3. Sort 200 mixed PDFs in `~/Downloads/` into year/month subfolders, generate an Excel index, and produce a one-page summary
4. Brainstorm taglines with a colleague over a chat thread

In the Cowork task loop, your role is primarily…?

1. Hands-off — kick off the task and let Claude run end-to-end
2. Continuous co-pilot — type alongside Claude on every step
3. Plan reviewer + outcome auditor — describe the goal, approve plans, sample-check results, halt early when direction is wrong
4. Just a log reader — Cowork can't be interrupted once running

A Cowork Project differs from a one-shot Cowork task in that it…?

1. Runs in Anthropic's cloud, freeing your laptop
2. Persists instructions, file context, scheduled tasks, and memory inside a named local workspace — repeatable work without re-explaining context
3. Is the same concept as a Claude Code project; just a UI rename
4. Makes the Project's data shared across everyone in your team automatically

You set up "every Friday 5 pm: generate weekly digest." On Friday at 5 pm your laptop is closed in your bag. What happens?

1. Cowork runs anyway in Anthropic's cloud — output appears Monday
2. Nothing — Cowork scheduled tasks need Claude Desktop running and the computer awake; missed runs may catch up next time you open the app, depending on the schedule
3. Your phone runs it via push notification
4. The next launch raises an error and disables the schedule

Inside Cowork, what's the relationship between Plugins, Skills, Connectors, and Sub-agents?

1. They're four names for the same connector concept
2. Skills are the umbrella; the others plug into a Skill
3. A Plugin packages Skills (how to do something), Connectors (where to reach), and Sub-agents (who handles which part) into a specialist workflow for a role
4. Sub-agents only exist in Claude Code, not in Cowork

When Claude generates text, what is it actually predicting?

1. The semantic meaning of the next sentence as a whole
2. The full final answer, then it checks back from the end
3. The next token, given all prior tokens — repeated until a stop condition
4. A vector embedding that's later decoded into characters

Claude confidently states a fact that turns out to be wrong. The most accurate explanation?

1. The model is broken; report it for retraining
2. Hallucination is an inherent risk of next-token prediction over imperfect training data — verification, citations, and RAG are the mitigations, not "telling Claude to be careful"
3. Claude is intentionally deceptive in some scenarios
4. It happens only in low-temperature settings

Pre-training, fine-tuning, and RLHF — pick the order and what each contributes.

1. RLHF → fine-tuning → pre-training, building from style to knowledge
2. They are alternatives; you pick one based on use case
3. Pre-training (next-token on huge corpora — broad world knowledge) → fine-tuning (task-specific data — sharper behavior) → RLHF (human preference signals — alignment with helpfulness/safety)
4. Only pre-training matters; the rest are marketing names

Today's date is past Claude's training cutoff. Which question is the riskiest to take Claude's answer at face value?

1. "Explain the origins of the Roman numeral system"
2. "Write a Python function to reverse a string"
3. "Who's currently the CEO of [recent-news company X]?"
4. "What's a metaphor for resilience?"

A teammate writes "make this email better" as a Claude prompt. Why does it underperform "rewrite this email to be 100 words, polite-firm tone, ending with a clear next-step request"?

1. Shorter prompts always work worse
2. It needs a higher temperature setting
3. "Better" is undefined; Claude has no shared rubric. Specifying length, tone, and required structure replaces ambiguous adjectives with measurable constraints
4. You must use XML tags or it won't work

Three colleagues describe Claude Code. Which one has the right product mental model?

1. "It's a chatbot in a browser tab where I paste code snippets"
2. "It's a terminal-native AI pair-programmer that operates inside my repository — reads files, proposes edits, runs commands"
3. "It's an autocomplete extension; it suggests one line at a time"
4. "It's a no-code visual flow builder"

First time setting up Claude Code in a repo. What's the right first move?

1. Just run `claude` and start asking questions — it figures everything out
2. Run `/init` (or write a small `CLAUDE.md`) capturing the repo's stack, conventions, key directories, and "do not touch" zones — Claude loads this every session
3. Copy the entire codebase into the prompt to "give context"
4. Disable file write so Claude can only read

Claude proposes a 12-file edit. You glance, think it looks good, and want to accept. Best practice?

1. Accept all — Claude wouldn't propose something wrong
2. Reject everything — too risky to read
3. Review the diff, accept partial / file-by-file, run tests, and commit only what passed — Claude Code is a proposer, not an autopilot
4. Always run `git add -A` before reading the diff to "checkpoint"

When is Claude Code the WRONG tool?

1. You need to refactor a 30-file module
2. You want to add tests for a new feature
3. You need to brainstorm a product roadmap with a stakeholder over coffee — there's no codebase context, just a conversation
4. You want to debug a flaky CI job

Mid-task you realize Claude Code is heading the wrong direction. Best action?

1. Wait for it to finish; then start over
2. Force-kill the terminal and discard everything
3. Press Esc / interrupt, redirect with new context, and use `/clear` only if the off-track context is irreversible — keeping useful history saves restart cost
4. Open a separate `claude` session in another terminal so they vote

In the 4D framework (Delegation / Description / Discernment / Diligence), which D handles "deciding whether AI should do this task at all"?

1. Description
2. Delegation
3. Discernment
4. Diligence

Your nonprofit uses Claude to research grant opportunities. The output lists 5 promising funders. Discernment step says you should…?

1. Trust the list — Claude has access to current grant databases
2. Verify each funder still exists, the deadline is current, and the eligibility criteria match — Claude may hallucinate or use outdated data
3. Ask Claude to verify itself by re-running the prompt
4. Forward the list to staff without review

Which type of nonprofit data is the LEAST safe to paste into a public Claude conversation?

1. Your annual report PDF (already public)
2. Generic fundraising email templates
3. Beneficiary case files containing names, contact info, health details, or immigration status
4. Your mission statement

"Diligence" in the 4D framework operationally means…?

1. Working harder than the AI does
2. Owning the final outcome — humans remain accountable for what was sent / decided / published, and document the AI's role + limits in their workflow
3. Reading every Claude transcript end-to-end
4. Running the same prompt three times for consensus

Your director asks Claude to analyze 5 years of donation data and recommend a fundraising strategy. When is AI-only analysis NOT enough?

1. Always — AI can never do data analysis
2. Never — AI handles it end-to-end if you upload the spreadsheet
3. When decisions affect real budgets, staff, or beneficiaries — AI surfaces patterns and questions, humans validate the math, judge the strategy, and own the call
4. Only when the dataset is over 100k rows

You're building a chat app that talks to Claude. Where should you put your API key?

1. In the user's browser
2. On your secure server
3. In your website's JavaScript code
4. In a public GitHub repository

What is the primary purpose of a system prompt when working with Claude?

1. To authenticate API requests to the Anthropic service
2. To provide instructions that customize Claude's tone, style, and approach
3. To limit the number of tokens Claude can generate in a response
4. To store the conversation history between multiple requests

Your users complain that your chat app feels slow because they wait 20 seconds staring at a loading spinner, then a bunch generated text suddenly appears. What feature should you add to fix this?

1. Shorter prompts
2. Response streaming
3. Multiple chatbots
4. Faster internet connection

You want Claude to generate only clean JSON code without any explanations or markdown formatting. Which combination of techniques works best?

1. Request shorter responses
2. Ask nicely in the prompt
3. Use high temperature and long prompts
4. Prefill with "{" and use "```" as a stop sequence

You're building a chatbot to answer factual questions about your company. You want consistent, reliable answers every time. What temperature setting should you use?

1. 0.1 (low temperature)
2. It doesn't matter
3. 0.8 (high temperature)
4. 1.0 (high temperature)

Claude reads your message "I love quantum physics." What happens first?

1. Claude breaks the text into smaller pieces called tokens
2. Claude researches quantum physics
3. Claude writes a response immediately
4. Claude translates it to another language

You ask Claude "What's the best programming language?" but you want it to specifically argue for Python. What technique helps you control this?

1. Use a higher temperature setting
2. Ask the question multiple times
3. Make the text bigger
4. Add an assistant message starting with "Python is the best because"

You're giving your AI a long document to analyze along with your instructions. What would help the AI understand your prompt better?

1. Put everything in one big paragraph
2. Write the document in all capital letters
3. Use XML tags like <document> and <instructions>
4. Separate sections with lots of blank lines

You want your AI to write a book review. Which approach would be more helpful?

1. "Write a book review that's 300 words, includes plot summary, mentions two characters, and gives a rating"
2. "Write a book review that's good and interesting"
3. "Tell me what you think about books in general"
4. "Write something about the book I just read"

You want to improve a prompt that isn't working well. What should you do first after writing your initial prompt?

1. Test it and measure how well it performs
2. Add more examples immediately
3. Rewrite it completely from scratch
4. Make it longer and more detailed

Your evaluation report shows your prompt scored poorly on "missing calorie information" across multiple test cases. What does this tell you?

1. You should ignore this feedback and try something else
2. You need to specifically tell your prompt to include calorie information
3. The test cases are too hard
4. The evaluation system is broken

You want your AI to detect sarcasm in social media posts, but it keeps missing sarcastic comments. What would help most?

1. Tell it to "try harder" to find sarcasm
2. Use bigger fonts in your prompt
3. Show it examples of sarcastic posts with correct labels
4. Ask it to guess when posts might be sarcastic

Why is the first line of your prompt considered the most important part?

1. It determines which AI model will be used
2. It determines how fast the AI will respond
3. It sets the stage for everything that follows and should be clear and direct
4. It controls the length of the AI's response

You've learned techniques for writing better prompts, but now you want to measure how well they actually work. What do you need?

1. More prompt engineering techniques
2. Prompt evaluation methods
3. More training data
4. A faster AI model

You need test cases for your prompt evaluation but don't want to write them all by hand. What's a good alternative?

1. Ask users to create test cases
2. Only test with one example
3. Skip testing and deploy immediately
4. Use Claude to generate test cases automatically

You write a prompt and test it twice with your own inputs. It looks good, so you deploy it. What's the main risk?

1. The prompt will work too slowly
2. The prompt will become too expensive
3. The AI model will stop working
4. Users might provide unexpected inputs that break it

In a typical evaluation workflow, what happens right after you feed your prompts through Claude?

1. You change the prompt and start over
2. You feed the responses through a grader
3. You create a new dataset
4. You deploy the prompt to production

When Claude wants to use a tool, it sends back a response that's different from usual. What does this response contain?

1. Error messages only
2. Both text blocks and tool use blocks
3. Only tool requests with no text
4. Only text like normal

You want to create a tool that gets the current time. What type of code do you need to write?

1. A regular Python function
2. A web page
3. A database query
4. A complex AI algorithm

A user asks Claude "What day will it be 30 days from today?" To answer this, Claude needs to use multiple tools. What happens?

1. Claude asks the user to do the math
2. Claude uses one tool and guesses the rest
3. Claude calls tools in sequence — first getting today's date, then adding 30 days
4. Claude gives up and says it can't help

Sarah asks Claude "What's the weather like today?" but Claude says it doesn't have current weather data. What would solve this problem?

1. Waiting for Claude to update itself
2. Giving Claude access to tools that fetch current data
3. Asking Claude to guess the weather
4. Training Claude on more weather information

You're building a chat app with Claude. A user asks for today's stock prices, but Claude responds "I don't have access to current stock information." What's the core problem?

1. The user asked the wrong question
2. Claude is broken
3. Claude only knows information from its training data
4. Claude needs to be restarted

You want to give Claude the ability to search the web for current information. What do you need to implement?

1. Just a simple schema — Claude handles the searching
2. Your own search engine
3. A complex web scraping system
4. Permission from Google

You've written a Python function for Claude to use. What else do you need so Claude knows how to call it?

1. Permission from Claude
2. A JSON schema describing the function
3. A special license
4. A user manual

You're setting up a system to handle large documents. Instead of using everything at once, you break documents into smaller pieces and search for relevant ones. What is this approach called?

1. The chunking approach
2. File splitting
3. Text summarization
4. Document compression

You try to include a massive 800-page document directly in your Claude prompt. What problems will you likely face?

1. There are hard limits on text length, reduced effectiveness, and higher costs
2. The document will be perfectly processed
3. Claude will work faster than normal
4. Only the cost will increase slightly

You have an 800-page financial report and want to ask Claude specific questions about it. What does RAG help you do?

1. Ask only yes/no questions
2. Put the entire document into each prompt
3. Summarize the whole document first
4. Find and include only the relevant sections for each question

You send the text "The cat is happy" to an embedding model. What do you get back?

1. A summary of the text
2. A translation in another language
3. A list of keywords
4. A long list of numbers

What problem does contextual retrieval solve in RAG systems?

1. It makes search queries run faster
2. It reduces the storage space needed for embeddings
3. It addresses the issue of chunks losing their connection to broader document context when documents are split
4. It eliminates the need for vector databases

You have search results from both semantic search and BM25 search. They use different scoring systems. How do you combine them into one ranked list?

1. Use Reciprocal Rank Fusion (RRF) based on rank positions
2. Take the average of both scores
3. Add the scores together directly
4. Use only the semantic search results

What is the purpose of re-ranking in RAG pipelines?

1. To compress the vector database for faster searches
2. To generate better embeddings for text chunks
3. To use an LLM to intelligently reorder search results after initial retrieval
4. To split documents into more appropriate chunk sizes

You're searching for a specific incident ID like "INC-2023-Q4-011" in your documents. Semantic search isn't finding it well. What additional search method would help?

1. Bigger vector database
2. BM25 lexical search for exact term matching
3. Longer embeddings
4. More chunks

When Extended Thinking is enabled, what two parts will Claude's response contain?

1. A summary block and a detail block
2. A thinking block and a text block
3. A draft block and a final block
4. A question block and an answer block

You ask Claude "How many marbles are in this image?" but get the wrong count. What's the best way to improve accuracy?

1. Ask the question in all capital letters
2. Send a higher quality image
3. Upload the image multiple times
4. Provide detailed counting steps and methodology

What's the minimum amount of content needed for caching to work?

1. Any amount of text
2. 500 tokens
3. 1024 tokens
4. 2000 tokens

You want to cache your tool definitions. Where should you place the cache breakpoint?

1. On the last tool in your list
2. On the middle tool in your list
3. On every tool in your list
4. On the first tool in your list

How does prompt caching work?

1. It makes Claude remember conversations forever
2. It prevents Claude from making mistakes
3. It reuses computational work from previous requests
4. It translates messages into different languages

You're building an app where users ask questions about documents. What's the main benefit of enabling citations?

1. It reduces the cost of each request
2. It shows users exactly where information came from
3. It makes Claude's responses longer
4. It makes the app run faster

User-controlled workflows that are triggered through UI interactions like button clicks or slash commands. This definition describes which MCP primitive?

1. Resources
2. Sessions
3. Tools
4. Prompts

What does "transport agnostic" mean in the context of MCP communication?

1. MCP automatically chooses the fastest available network connection
2. MCP requires specific hardware to function properly
3. MCP only works with HTTP connections
4. MCP clients and servers can communicate using different methods like HTTP, WebSockets, or standard input/output

What are Resources in the context of MCP?

1. App-controlled data access for UI purposes or adding context to conversations
2. Model-controlled functions for performing calculations
3. User-triggered commands that start predefined workflows
4. Server configuration settings that control performance

In MCP architecture, what is the relationship between MCP Clients and MCP Servers?

1. MCP Clients generate AI responses while MCP Servers handle user input
2. MCP Clients store data while MCP Servers process requests
3. MCP Clients connect to MCP Servers that contain tools, prompts, and resources
4. MCP Clients and MCP Servers are the same component with different names

What is Model Context Protocol (MCP)?

1. A programming language specifically designed for AI applications
2. A communication layer that provides Claude with context and tools without requiring tedious integration code
3. A security protocol for encrypting AI model responses
4. A database management system for storing AI conversations

What is the MCP Server Inspector?

1. A command-line tool for monitoring server performance
2. A code editor specifically designed for writing MCP servers
3. A security tool for scanning MCP servers for vulnerabilities
4. A browser-based interface for testing and debugging MCP servers in real-time

Which of the following correctly describes Tools in MCP?

1. App-controlled data that populates UI elements
2. Static configuration files that define server behavior
3. User-controlled workflows that can be triggered on demand
4. Model-controlled functions that Claude decides when to call

You want Claude to analyze a product image for 6 different materials at once. Instead of one huge prompt, what should you do?

1. Write a longer, more detailed prompt
2. Send 6 separate requests in parallel, then combine results
3. Ask the user to pick one material first
4. Use an agent with material tools

You're building a system where Claude creates content, then checks if it's good enough, then improves it if needed. What pattern is this?

1. Parallelization
2. Evaluator-optimizer
3. Routing
4. Chaining

You're building an app where users upload photos and always get the same 4-step process to enhance them. Which approach should you use?

1. An agent with photo tools
2. A single complex prompt
3. Multiple agents working together
4. A workflow with predefined steps

Your app needs to handle both "cooking recipes" and "workout routines" with completely different styles. What pattern helps?

1. Combine both styles in one prompt
2. Use the same prompt for both
3. Routing — categorize first, then use specialized prompts
4. Always ask users to specify the category

What defines an agent when working with Claude?

1. A predetermined sequence of steps that Claude must follow exactly
2. A setup where Claude is given a goal and tools, then figures out how to complete the goal
3. A system that categorizes user requests into different types
4. A method for breaking complex tasks into parallel subtasks

What is environment inspection in the context of AI agents?

1. The process of categorizing user requests into different workflow types
2. The technique of running multiple specialized tasks simultaneously
3. A method for breaking large tasks into smaller sequential steps
4. Claude's ability to observe and understand the results of its actions

You're building an agent. In general, should you give it a "Refactor Code" tool or basic tools like "read file" and "write file"?

1. "Refactor Code" tool — it's more specific
2. Basic tools — they're more flexible
3. Both tools together
4. Neither — agents don't need tools

You're porting a Python script from `anthropic.Anthropic()` (direct API) to `anthropic.AnthropicVertex(...)`. The `model` parameter must change how?

1. It stays the same — `claude-sonnet-4-5`
2. Prefix with `vertex.`, e.g. `vertex.claude-sonnet-4-5`
3. Use the Vertex Model Garden ID, typically `claude-sonnet-4-5@` or the current endpoint alias
4. The SDK rewrites it automatically; pass anything

Most of your users are in the US and you need predictable latency plus US-only data residency. Endpoint choice?

1. `global` — let Vertex route freely
2. A US multi-region endpoint (`us`) or a specific US region — keeps requests within US borders
3. `eu` — closer to global average
4. Skip Vertex; call Anthropic API directly for lower latency

Production auth on Vertex — best pattern when running on GKE?

1. Hard-code Service Account JSON in container image
2. Reuse the existing `ANTHROPIC_API_KEY` env var
3. Workload Identity — Pod assumes a Google Service Account without any JSON key on disk
4. Have users sign in with their personal Google accounts

Compliance auditor asks "show me every Claude call your team made last month, with caller identity." Where do you look?

1. Anthropic Console usage page
2. Cloud Audit Logs in the GCP project — Vertex records every API call with caller identity and metadata
3. Application logs only — Vertex doesn't expose this
4. It's not retained; you'd have to log it yourself manually going forward

If you write Vertex requests with the native REST body (not via the Anthropic SDK), what's the gotcha?

1. Native REST is forbidden; you must use the SDK
2. You must include `anthropic_version: "vertex-2023-10-16"` in the request body — Vertex requires it
3. The body is identical to Anthropic API requests, no changes
4. You must encrypt the body with a public key first

Migrating an Anthropic-API-built app to Vertex. Which step is most often skipped and most likely to break production?

1. Renaming the GCP project
2. Rotating the (no-longer-used) Anthropic API key
3. Re-running your eval set against the Vertex-versioned model and target region — quotas, latencies, and (rarely) behavior nuances differ
4. Switching all calls to Sonnet

You're building a chat app and users complain that responses take too long to appear. What feature should you implement?

1. Send requests faster
2. Add a loading spinner
3. Make the messages shorter
4. Use streaming to show text as it's generated

You're using Claude to extract data from documents and need the same consistent format every time. What temperature setting should you use?

1. Temperature close to 0 for consistent, predictable outputs
2. Temperature 0.5 for balanced responses
3. Temperature 1.0 for maximum creativity
4. Temperature doesn't matter for data extraction

You want to build a customer service bot that only talks about your company's products and stays professional. What's the best approach?

1. Tell users to only ask product questions
2. Add "be professional" to every user message
3. Use a system prompt that makes Claude act like a customer service representative
4. Set the temperature to maximum creativity

You're building a chatbot where users ask follow-up questions. A user asks "What's 2+2?" and then asks "Add 5 to that." What do you need to do for the second question to make sense?

1. Send both the first question and Claude's previous answer along with the second question
2. Restart the conversation from the beginning
3. Wait 30 seconds before sending the second question
4. Send only the second question to Claude

You want an AI to write movie reviews in a specific style. What's the best way to show the AI exactly what you want?

1. Use very strict formatting rules
2. Tell it to copy famous movie critics
3. Describe the style in great detail
4. Give it a sample movie review as an example

You're improving a prompt that generates workout plans. What should you do after writing your first version?

1. Test it, see how well it works, then improve it
2. Use it immediately for all workouts
3. Write five different versions at once
4. Ask other people to guess what it does

You're asking an AI to analyze a long customer review mixed in with your instructions. What helps the AI understand which part is the review?

1. Put the review at the very end
2. Put the review between XML tags like <review></review>
3. Write the review in a different font
4. Make the review all uppercase

You want an AI to write a book summary. Which opening instruction works best?

1. Write a three-paragraph summary of this book
2. What do you think about summarizing things?
3. I was wondering if you could maybe help with something about books?
4. Books are interesting, aren't they?

You need test data for evaluating your prompt. You want to create realistic examples quickly without writing them all by hand. What's the best approach?

1. Use the same example over and over
2. Use Claude to automatically generate test cases
3. Ask your friends to write them
4. Copy examples from the internet

You've written a prompt for Claude and want to know if it works well. What's the difference between prompt engineering and prompt evaluation?

1. Prompt engineering is for beginners, prompt evaluation is for experts
2. Prompt engineering tests the prompt, prompt evaluation writes it
3. Prompt engineering writes better prompts, prompt evaluation measures how well they work
4. They're the same thing with different names

In prompt evaluation, what is a "grader" used for?

1. To save your prompts to a file
2. To give objective scores measuring output quality
3. To write better prompts automatically
4. To make Claude respond faster

You just wrote a prompt for your app. You test it once and it works great, so you decide to use it. What's the main risk with this approach?

1. The prompt will stop working after a few days
2. Users will provide unexpected inputs that break it
3. It will be too expensive to run
4. Other developers won't understand your code

You're running a prompt evaluation. After creating your dataset and feeding questions through Claude, what's the next step?

1. Feed the responses through a grader to get scores
2. Write a completely new prompt
3. Ask users what they think
4. Publish your prompt immediately

You're using another AI model to evaluate Claude's responses. To get better scores than just random numbers around 6, what should you ask the grader to provide?

1. Just "good" or "bad"
2. Strengths, weaknesses, reasoning, and a score
3. A rewritten version of the response
4. Only a number from 1-10

You want to check if Claude's output contains certain keywords and has the right length. Which type of grader should you use?

1. Manual grader
2. Code grader
3. Model grader
4. Human grader

When using tools, what happens right after Claude asks for specific external data?

1. Claude analyzes the original question again
2. The user needs to approve the data request
3. Your server runs code to fetch the requested information
4. Claude provides the final answer immediately

You want to force Claude to use a specific tool for data extraction. Which toolChoice setting should you use?

1. {"toolChoice": {"auto": {}}}
2. {"toolChoice": {"tool": {"name": "tool-name"}}}
3. {"toolChoice": {"any": {}}}
4. {"toolChoice": {"required": true}}

You ask Claude "What's the weather today?" but it says it doesn't have current weather information. What would tools help Claude do?

1. Access live weather data from external sources
2. Ask you to check the weather yourself
3. Guess the weather based on the date
4. Remember previous weather conversations

You're writing a tool function for Claude. What's the most important thing to include when creating the JSON schema?

1. The programming language being used
2. Detailed descriptions of what the tool does and its parameters
3. Your contact information
4. The function's source code

What is contextual retrieval?

1. A technique that adds context to document chunks before storing them to improve search accuracy
2. A way to reduce the size of document chunks
3. A system for automatically generating new content from existing documents
4. A method for searching through documents faster

What is a vector database in the context of RAG systems?

1. A specialized database optimized for storing, comparing, and searching through numerical embeddings
2. A system for backing up document files
3. A regular database that stores text documents as files
4. A database that only stores mathematical equations

You're searching for a specific incident ID "INC-2023-Q4-011" in your documents. Semantic search isn't finding it well. What search method would work better?

1. Converting everything to lowercase first
2. Searching only document titles
3. BM25 lexical search for exact keyword matching
4. Using longer text embeddings

You send the text "The cat sat on the mat" to an embedding model. What do you get back?

1. A shorter version of the same text
2. Keywords extracted from the text
3. A list of about 1024 numbers representing the meaning
4. A translation in another language

You have an 800-page financial report and want to ask an AI specific questions about it. What does RAG help you do?

1. Send only the relevant sections to the AI for each question
2. Make the document shorter by deleting pages
3. Translate the document into simpler language
4. Create a summary of the entire document

You send Claude the same long document twice in a row. What does prompt caching help with?

1. It stores your conversation history permanently
2. It automatically summarizes repeated content
3. It reduces the document's file size
4. It saves the computational work from processing the text in the document

You've optimized your prompt but Claude still isn't accurate enough on a complex task. What should you consider next?

1. Rewrite the prompt completely from scratch
2. Break the task into smaller pieces
3. Use extended thinking to improve accuracy
4. Switch to a different AI model

You want to cache a short message that's 500 tokens long. What will happen?

1. It will be automatically expanded to meet requirements
2. It won't be cached because it's too short
3. It will be cached normally
4. It will be cached at half price

What is an effective technique for increasing Claude's effectiveness with images?

1. Uploading more images
2. Using prompt engineering techniques
3. Using JPEG instead of PNG images
4. Providing zoomed-in images

You're building a chat app where users ask Claude about their GitHub data. Without MCP, what's the main problem you'd face?

1. Claude can't connect to the internet
2. GitHub doesn't allow API access
3. You'd have to write and maintain all the GitHub tool functions yourself
4. Users can't type GitHub questions

You're creating an MCP server tool using the Python SDK. What's the easiest way to define a new tool?

1. Create a separate configuration file
2. Send HTTP requests to register tools
3. Write complex JSON schemas manually
4. Use the @mcp.tool decorator on a function

Claude automatically decides to use a calculator tool when you ask "What's 15 × 23?" Who is controlling this tool usage?

1. The MCP server providing the tool
2. The application showing the chat
3. Claude (the AI model) itself
4. The user who asked the question

You just wrote an MCP server and want to test if your tools work correctly. What's the best first step?

1. Ask other developers to try it
2. Connect it to Claude immediately
3. Write unit tests for each function
4. Use the MCP Inspector in your browser

You want to let users type "@document_name" to automatically include document content in their message. Should you use a tool or a resource?

1. Tool - because Claude needs the information
2. Resource - because documents are files
3. Resource - because the app fetches data for the UI
4. Tool - because it involves documents

You're running an MCP client and server on the same computer. How do they most commonly communicate with each other?

1. Using Bluetooth connection
2. Through standard input/output
3. Through email messages
4. By writing files to disk

Your MCP client needs to find out what capabilities an MCP server offers. What message type should it send?

1. GetServerInfo
2. CheckCapabilities
3. ListToolsRequest
4. CallToolRequest

Your MCP server has tools, resources, and prompts. A user clicks a "Format Document" button in your app. Which primitive is being used?

1. Resources - because it accesses documents
2. Tools - because it formats something
3. All three at the same time
4. Prompts - because the user directly triggered it

Porting from Anthropic API to Bedrock. The model identifier changes how?

1. Stays the same — `claude-sonnet-4-5`
2. Bedrock model ID prefixes the family with `anthropic.`, e.g. `anthropic.claude-sonnet-4-5-v2:0` — and may need an inference profile prefix for cross-region capacity
3. Use the Vertex format `claude-sonnet-4-5@`
4. Bedrock auto-resolves any string

Why does your first call to a new Claude model on Bedrock fail with `AccessDeniedException`?

1. Bedrock requires an Anthropic API key as fallback
2. You need to email AWS support for every model
3. Bedrock requires you to request and grant model access in the Bedrock console before invocation; access is not auto-granted per account
4. The model is unavailable in any AWS region

Production app on EC2 calling Bedrock. Best auth pattern?

1. Hard-code AWS access key + secret in the app config
2. Reuse the Anthropic API key as `AWS_ACCESS_KEY_ID`
3. Attach an IAM role to the EC2 instance profile; SDK uses the role's temporary credentials
4. Use the AWS root account credentials

Cross-region inference profile — what does it actually do?

1. Reduces cost by 50% for the first 1k requests
2. Routes invocations across multiple AWS regions for capacity & availability without code changes
3. Replaces VPC endpoints
4. Required to enable streaming

Compliance team needs evidence of every Claude invocation. Where on AWS?

1. Bedrock console "History" tab — only shows the last 24h
2. CloudTrail records all Bedrock control-plane and (with Data Events enabled) data-plane API calls — including `InvokeModel` and `Converse`
3. Anthropic Console — Bedrock proxies traffic there
4. It's not captured; deploy your own logging

Migrating an Anthropic-API-built app to Bedrock. Most subtle compatibility risk to test for?

1. System prompts are removed in Bedrock
2. `max_tokens` is unsupported in Bedrock
3. Tool use, content block, and request/response shapes have minor differences — eval set must re-pass on the Bedrock SDK end-to-end
4. Streaming is unavailable on Bedrock

子智能体（subagent）与主会话最核心的区别是什么？

1. 子智能体只能运行 Bash 命令，不能编辑文件
2. 子智能体只能通过 slash command 手动调用
3. 子智能体拥有独立的上下文、系统提示词和工具权限，与主会话隔离
4. 子智能体的响应速度始终比主会话慢

以下哪项任务最适合委派给子智能体，而非在主会话中处理？

1. 修改单个配置文件中的一个变量名
2. 对全仓库进行安全扫描，只需要最终的问题清单
3. 与用户持续讨论架构方案，需要多轮交互
4. 需要密切参考主会话中刚刚讨论的上下文来做决策

子智能体配置中，哪两个字段是必填的？

1. name 和 model
2. name 和 tools
3. name 和 description
4. description 和 model

子智能体配置文件存放在 `.claude/agents/`（项目级）和 `~/.claude/agents/`（用户级）。以下说法正确的是？

1. 用户级子智能体优先级始终高于项目级
2. 项目级子智能体不能被团队成员使用
3. 项目级子智能体可随仓库提交，团队成员 clone 后即可使用
4. 两种级别的子智能体功能不同——用户级不支持 tool use

子智能体的 `description` 字段写为 "Helps with coding tasks"，这有什么问题？

1. description 太长，会浪费上下文
2. 太宽泛——几乎任何任务都可能触发，导致 Claude 误判委派时机
3. description 不能包含英文，必须用中文
4. description 中不能出现 "coding" 等通用词

为代码审查（code review）子智能体配置工具权限时，最佳实践是什么？

1. 赋予全部工具权限，以便子智能体可以自主修复发现的问题
2. 仅赋予 Read、Glob、Grep 等只读工具——审查不应修改代码
3. 赋予 Edit 和 Write 权限但禁止 Bash
4. 不配置 tools 字段，让子智能体继承主会话的完整权限

设计子智能体时，单一职责原则（Single Responsibility Principle）的核心含义是？

1. 每个子智能体只能使用一种工具
2. 每个子智能体应专注于一类明确的任务，而非充当万能助手
3. 每个项目中最多只能创建一个子智能体
4. 子智能体一次只能处理一个文件

关于子智能体的前台（foreground）与后台（background）执行，以下哪种场景最适合后台执行？

1. 代码修改任务，需要用户的权限确认
2. 代码审查，结果需要立即讨论
3. 对多个外部文档进行长时间研究搜索，结果供后续参考
4. 执行数据库迁移，需要确认迁移结果成功才能继续

一个 Agent Skill 的入口文件是什么？

1. README.md
2. AGENT.md
3. SKILL.md
4. index.md

Skill 的 `description` 字段主要作用是什么？

1. 在 UI 中展示 Skill 的版本信息
2. 帮助 Claude 判断何时自动调用该 Skill
3. 定义 Skill 可以使用的工具列表
4. 指定 Skill 的输出文件格式

项目级 Skill 的正确存储路径是什么？

1. ~/.claude/agents/<skill-name>/SKILL.md
2. .claude/skills/<skill-name>/SKILL.md
3. .claude/commands/<skill-name>/SKILL.md
4. skills/<skill-name>/SKILL.md

Skill 的 `allowed-tools` 字段的正确理解是什么？

1. 它定义了 Skill 被禁止使用的工具列表
2. 它预批准（pre-approve）所列工具——Skill 激活时这些工具无需用户额外确认
3. 它限制 Skill 只能使用这些工具，其他工具被完全禁止
4. 它继承自父智能体的工具权限，不可单独配置

一个 Skill 执行时会产生副作用（如发送 Slack 消息、创建 PR），最佳实践是什么？

1. 在 description 中写明 "This skill has side effects"
2. 不配置 allowed-tools，让用户逐次确认
3. 设置 disable-model-invocation: true，只允许用户手动调用
4. 将 Skill 放在用户级路径避免被团队误用

某团队需要给 Claude 接入公司内部数据库以实时查询客户信息，应选择哪种机制？

1. 创建一个包含数据库查询步骤的 Skill
2. 构建 MCP Server 暴露数据库查询工具
3. 在 CLAUDE.md 中写入数据库 schema
4. 创建一个包含数据库凭据的子智能体

以下哪种方式是团队共享 Skill 的主要途径？

1. 通过邮件发送 SKILL.md 文件给团队成员
2. 将 Skill 放在项目 .claude/skills/ 目录，随 git 提交
3. 上传到 Anthropic 官方 Skill 市场
4. 使用 ~/.claude/skills/ 路径并导出环境变量

Skill 无法被 Claude 自动发现，排查的第一步应该检查什么？

1. 文件路径是否正确，入口文件是否命名为 SKILL.md，frontmatter 是否有效
2. description 是否包含足够多的关键词
3. Skill 目录中是否包含 examples 文件夹
4. CLAUDE.md 中是否声明了该 Skill

AI Fluency 的四个目标是什么？

1. 快速、准确、低成本、可扩展
2. 有效（effective）、高效（efficient）、合乎伦理（ethical）、安全（safe）
3. 自动化、增强、代理、协作
4. 委派、描述、辨别、勤勉

「人类设定目标、边界和行为规则，AI 更独立地工作」描述的是哪种协作模式？

1. Automation（自动化）
2. Augmentation（增强）
3. Agency（代理）
4. Assistance（辅助）

4D 框架中的四个能力分别是什么？

1. Define, Design, Develop, Deploy
2. Discover, Diagnose, Decide, Deliver
3. Delegation, Description, Discernment, Diligence
4. Data, Dialogue, Decision, Deployment

Delegation（委派）的三要素不包括以下哪项？

1. Problem Awareness（问题意识）：我真正要达成什么？
2. Platform Awareness（平台意识）：AI 擅长和不擅长什么？
3. Task Delegation（任务委派）：哪些由人做、哪些由 AI 做？
4. Performance Evaluation（绩效评估）：AI 的输出效率如何？

Description（描述）的三种类型是什么？

1. 任务描述、格式描述、风格描述
2. Product Description（产品描述）、Process Description（过程描述）、Performance Description（表现描述）
3. 输入描述、处理描述、输出描述
4. 目标描述、方法描述、结果描述

Discernment（辨别）评估后，以下哪项不是正确的后续行动？

1. 直接采用（低风险格式化输出）
2. 要求修改（给出具体反馈）
3. 返回重新描述或切换工具
4. 忽略发现的问题，因为 AI 输出通常都会有小错

Description-Discernment 循环的核心逻辑是什么？

1. 写好提示词 → AI 生成 → 如果不好就换一个模型
2. 描述目标 → AI 生成/行动 → 评估产品/过程/表现 → 定位差距 → 调整描述或重新委派
3. 描述任务 → AI 自动评估 → 人类最终确认
4. 一次性写好详细提示词，避免反复修改

Diligence（勤勉）的三种类型是？

1. 数据勤勉、代码勤勉、内容勤勉
2. 输入勤勉、处理勤勉、输出勤勉
3. Creation Diligence（创建勤勉）、Transparency Diligence（透明勤勉）、Deployment Diligence（部署勤勉）
4. 事前勤勉、事中勤勉、事后勤勉

以下关于大语言模型「幻觉」（hallucination）的说法，哪项最准确？

1. 大模型不会产生幻觉，只要使用最新版本即可避免
2. AI 输出可能流畅自信但包含虚构事实，因为模型基于模式生成而非从数据库检索
3. 幻觉只会出现在中文等非英语语言中
4. 只要在系统提示词中写「不要编造」，就可以杜绝幻觉

以下哪项属于 Diligence 的常见失败模式？

1. 使用 AI 生成初稿后人工修改
2. 将未核实的 AI 输出作为事实直接发布
3. 在报告中声明使用了 AI 辅助
4. 根据任务风险等级决定 AI 的参与程度

教育场景中 AI 应用的四个层次不包括以下哪项？

1. 课程层（Course layer）：教学大纲、活动设计
2. 材料层（Material layer）：阅读材料、示例、练习题
3. 管理层层（Administration layer）：排课、考勤、预算
4. 制度层（Institutional layer）：政策草案、伦理规范

设计课堂 AI 活动的五个步骤中，「分配辨别任务」的目的是什么？

1. 让学生比较不同 AI 工具的速度
2. 让学生主动评估 AI 输出的准确性、偏见和适用性，培养批判性思维
3. 让教师检查 AI 是否按预期工作
4. 为 AI 输出打分，确定哪些学生可以用 AI

在 AI 时代，学术诚信的评估框架发生了什么转变？

1. 从「是否注明引用来源」转变为「是否使用了 AI」
2. 从「这是你自己的作品吗」转变为「使用方式是否促进了真正的学习」
3. 从「是否抄袭」转变为「AI 使用比例是否低于 30%」
4. 从「独立完成」转变为「鼓励使用 AI 完成全部作业」

以下哪个学科中 AI 的「辨别需求」最高——即最需要人工核验 AI 输出？

1. 创意写作——AI 可以提供风格建议
2. 法律——AI 可能生成看似合理但引用错误的法律条文
3. 编程入门——AI 可以解释语法错误
4. 体育——AI 可以分析运动数据

教育者视角的 Delegation-Diligence 循环中，第一步应该做什么？

1. 让 AI 生成教学材料的第一版
2. 教师先确定哪些任务不可委派给 AI（如学习目标设定、最终评分）
3. 让学生试用 AI 工具并收集反馈
4. 查阅其他学校的 AI 使用政策作为模板

关于作业设计，材料中建议的核心理念是什么？

1. 完全禁止 AI 在作业中的使用，以保持学术诚信
2. 从「检测 AI 使用」转向「让学生展示理解过程、判断力和责任」
3. 允许学生在任何作业中自由使用 AI，无需声明
4. 为每份作业设置 AI 使用比例上限（如不超过 30%）

教授 AI Fluency 时，教学的核心应该是什么？

1. 教学生使用最新 AI 工具的具体按钮和功能
2. 教可迁移的判断框架，而非特定工具的操作技能
3. 教学生写出最优美的提示词模板
4. 教学生比较不同 AI 模型的性能基准

四种 AI Fluency 教学法中，哪种最接近真实 AI 协作场景但需要最多的课堂时间？

1. Linear（线性）：按顺序教四个 D
2. Non-linear（非线性）：根据任务需要跳跃式教学
3. Focused（聚焦）：在一个 D 上深度教学
4. Loop-based（循环式）：围绕两大循环设计活动

Delegation-Diligence 循环教学活动中，最重要的产出是什么？

1. 学生对分工判断和责任方式的清晰解释
2. AI 生成的高质量最终作品
3. 学生编写的精美提示词
4. AI 使用的时间效率数据

评量 AI Fluency 时，最稳健的策略是什么？

1. 仅看最终作品质量
2. 仅看学生的 AI 聊天记录
3. 结合结果（最终作品）、过程（版本变化、决策记录）和反思（学生自述），三者并用
4. 仅看学生对 AI 使用的自我反思

AI Fluency 作业设计的三个原则是什么？

1. 简单、快速、可自动评分
2. 真实性（Authenticity）、迭代（Iteration）、文档化（Documentation）
3. 标准化、统一化、集中化
4. 开放性、创造性、协作性

分析 AI 对学科的影响时，以下哪种判断是正确的？

1. 所有学科的课程内容都应保持不变，只需在考试中禁用 AI
2. 学科中容易被 AI 替代的操作需要更新教学内容，难以替代的判断力则更加珍贵
3. 所有学科的评估方式都应改为口试
4. AI 对 STEM 学科没有影响，只影响人文学科

学生在学习场景中使用 AI 时，最关键的边界是什么？

1. 每天使用 AI 的时间不超过 2 小时
2. 增强（Augmentation）vs 自动化（Automation）——AI 应让你更有能力，而非替你完成该学的内容
3. 只能使用学校指定的 AI 工具
4. 使用 AI 时必须全程录屏以便教师检查

「铁律」（Iron Rule）的核心含义是什么？

1. 永远不要使用 AI 完成任何作业
2. 无论 AI 参与多少，所有提交内容必须是你能够解释、应用和负责的
3. AI 只能用于课外学习，不能用于课堂相关任务
4. 每次使用 AI 后必须向老师提交使用报告

学生版 Learning Context Document（学习背景文档）不包含以下哪项？

1. 你是谁（年级、专业、当前课程）
2. 学习目标（这次学习要真正掌握什么）
3. 当前水平（已知什么、卡在哪里）
4. 过往所有考试成绩和 GPA

AI 作为学习伙伴时，以下哪种使用方式最有利于学习？

1. 让 AI 生成完整答案后抄写到作业本上
2. 让 AI 生成练习题（不含答案），自己先做，之后让 AI 给予反馈
3. 考试时用 AI 实时查询答案
4. 用 AI 代写读书笔记，节省时间多看几本书

关于使用 AI 准备求职材料，以下哪项是正确的？

1. AI 可以为没有实习经历的学生编造合理的项目经验
2. AI 可以帮助改善简历的表达清晰度，但真实经历和个人声音必须由你保持
3. 直接提交 AI 生成的通用求职信是最有效率的方式
4. AI 可以替学生决定最适合的职业方向

提交 AI 辅助作业前的最终自检（Self-Check）不包含以下哪项？

1. 我能在没有 AI 的情况下解释这个内容吗？
2. 我知道哪些部分来自 AI，哪些是我自己的判断吗？
3. 我核实了事实、来源或计算吗？
4. 我使用了最新的 AI 模型版本吗？