Practice · 28 questions

CCDV-F — Mock 1 · Comprehensive

Weight-proportional coverage across all eight domains — Claude API mechanics, model selection and cost, agents vs. workflows, prompt/context engineering, tools and MCP (including the July 2026 spec revision), security, Claude Code, and debugging. Every option is plausible and roughly the same length, so you have to reason, not pattern-match. Answer as many as you like (unanswered count as incorrect), then get the explanation for every choice.

Architect-level reasoning Distractors = real misconceptions Per-domain scorecard

How to use it · Commit to an answer before revealing the key. When you review, read the explanation for the options you didn't pick too — the distractors encode the exact misconceptions the real exam exploits. This is a study quiz, not the graded platform exam.

0 of 28 answered

V2 · Q1Streaming vs non-streaming for long-running requests
A team's backend calls the Messages API synchronously (no streaming) with a large max_tokens value for a report-generation feature. In production, some requests silently fail with no response after several minutes, even though smaller requests succeed reliably. What is the most likely cause and fix?
V2 · Q2Streaming event sequence and stop_reason
A developer building a chat UI wants to detect the moment a tool call's input is fully assembled during streaming, and separately wants to read why generation stopped, without waiting for a second request. Which SSE events should they rely on?
V2 · Q3Retryable vs permanent errors
An application's error handler currently retries every non-2xx response from the Messages API with exponential backoff, including 400s. During a load test, a bug causes every request to be malformed (missing max_tokens), and the retry loop hammers the API with the same broken request for minutes before the test times out. What should change?
V2 · Q4Idempotency and duplicate side effects on retry
An agent uses a custom tool that calls a third-party payment API. A network blip causes the Messages API request that triggered the tool call to time out; the SDK's default retry logic resends the request, and the tool ends up being executed twice, charging the customer twice. What's the correct diagnosis?
V2 · Q5Auth for production/CI workloads
A platform team is setting up CI pipelines that call the Claude API from short-lived build containers. They want to avoid storing any long-lived secret in the CI system, and want credentials to expire quickly if a container is ever compromised. Which authentication approach best fits this goal?
V2 · Q6Token-bucket rate limiting
A team notices their application gets occasional 429s even though their traffic is well below their plan's stated requests-per-minute limit when averaged over each calendar minute — bursts right at the start of a minute sometimes get throttled. What best explains this?
V2 · Q7Streaming vs Batch API tradeoffs
A data team needs to classify 200,000 support tickets overnight, with results needed by morning but no requirement to see any single result immediately. Which approach best fits Anthropic's guidance for this workload?
V2 · Q8Handling pause_turn in agentic loops
An agent using server-side tools (like web search) in a loop occasionally receives stop_reason: "pause_turn" instead of end_turn or tool_use. The harness currently treats any non-tool_use stop_reason as "done" and returns the response to the user. What's wrong with that, and what should happen instead?
V2 · Q9Secrets management and version control discipline
A developer wants to unblock local testing quickly and commits a working Claude API key directly into a feature branch's source file "just for now," planning to remove it before merging. A teammate flags this in code review. What best practice is being violated, and what should replace it?
V5 · Q10Model tier tradeoffs
A team is building a customer-support triage feature that classifies incoming tickets into one of 8 fixed categories, running on millions of tickets per month, with no complex reasoning required. Cost and throughput matter far more than nuanced judgment. Which model choice best fits this workload given the current tier lineup?
V5 · Q11Prompt caching mechanics and limits
A team adds a cache_control breakpoint at the end of their system prompt, expecting to cache a large shared instruction block. In production, cache_read_input_tokens stays at zero even though the same system prompt is sent on every request. They discover their system prompt is only about 700 tokens, and they're running an older model whose minimum cacheable prefix is 1,024 tokens. What's happening?
V5 · Q12Prompt caching TTL economics
A team caches a large tool-and-system prefix using the default 5-minute TTL, but their traffic pattern has long gaps — a burst of requests every 20 to 30 minutes, with just a few requests close together in each burst. They're seeing high cache-write costs relative to cache reads. What should they consider, and why?
V5 · Q13Token counting across model generations
A team maintains a cost-estimation dashboard that multiplies text length by a fixed tokens-per-character ratio they measured once against an older Sonnet model. After migrating some traffic to a newer-tokenizer model, the dashboard's estimates start under-predicting actual token usage and cost. What's the most likely explanation, and what should they do?
V5 · Q14Extended thinking mechanics across model tiers
A team has agentic code that sets thinking: {type: "enabled", budget_tokens: 8000} and it works fine when routed to claude-haiku-4-5. When they add a fallback path that reroutes the same request shape to claude-opus-5 during Haiku overload, the fallback path starts failing. What's the most likely cause?
V1 · Q15Workflow vs. agent selection
A fintech team is building a system that ingests a monthly expense report, extracts vendor and amount fields, checks the amount against a fixed threshold table, and files the report into one of three predetermined folders based on category. The steps and their order never change. A junior engineer proposes building this as an autonomous Claude agent that "decides for itself" how to process each report. What should the team do instead, per Anthropic's agents-vs-workflows framework?
V1 · Q16Evaluator-optimizer pattern
A team wants Claude to draft a legal disclaimer, have a second Claude call critique the draft against a style guide, and loop the draft-critique cycle using that feedback until the critique reports no more issues or a maximum of three iterations is reached. Which workflow pattern does this describe?
V1 · Q17Parallelization: sectioning vs. voting
A content safety team wants higher confidence on a borderline moderation decision, so they send the identical flagged post to Claude in five separate, simultaneous calls and take the majority verdict across the five responses. Which parallelization sub-pattern is this, and how does it differ from sectioning?
V1 · Q18Claude Agent SDK vs. Managed Agents hosting
A startup wants to ship an agent with the full Claude Code harness capabilities (context management, hooks, subagents, MCP, permissions) inside their own product, but they explicitly do not want to run or maintain any sandbox or execution infrastructure themselves. Which option fits this requirement, and why does the most obvious-sounding alternative fall short?
V8 · Q19Tool description quality
A developer defines a refund_order tool whose description only says "Issues a refund for an order." In practice, Claude frequently calls this tool for plain order-status questions where no refund should occur. What is the most likely fix, per Anthropic's tool-use guidance?
V8 · Q20tool_result and is_error handling
An app calls a lookup_account tool. The account ID Claude generated doesn't exist, so the app's tool-execution code throws. The developer returns a tool_result block with content: "Account not found" and the correct matching tool_use_id, but omits the is_error field. Claude subsequently treats the string as if it were valid account data. What is the fix?
V8 · Q21MCP spec revision (2026-07-28)
A team last reviewed the Model Context Protocol spec in 2025 and built a remote MCP server that pushes real-time progress notifications to clients by default and registers new clients using Dynamic Client Registration (DCR). After the 2026-07-28 MCP spec revision, which statement about their design is accurate?
V3 · Q22Claude Code settings precedence
An engineering org sets a permission policy through managed (MDM) settings that disallows bypassPermissions mode organization-wide. An individual engineer, wanting faster local iteration, adds "permissions.mode": "bypassPermissions" to their own .claude/settings.local.json. What happens when Claude Code resolves settings for that engineer's session?
V6 · Q23Context engineering: isolating large tool outputs with sub-agents
A coding agent needs to search a large monorepo for every usage of a deprecated API and then present the user with a short migration plan. The team is choosing between running the search directly in the main agent loop versus dispatching it to a sub-agent. They've noticed that as sessions run longer, the agent's reasoning quality visibly degrades. Which design best applies context engineering principles here?
V6 · Q24Prompt structure: the system parameter vs. messages, and caching stability
A team building against the Messages API wants to keep their integration simple, so they put all persona, formatting, and tool-use policy instructions as the first entry in the messages array (role: 'user'), followed by the actual per-request user content as later 'user' role messages. What should they change, and why?
V6 · Q25Structured output: syntactic validity vs. semantic correctness
A team uses a documented structured-output technique to get Claude to return JSON for a form-filling feature. During testing, every response parses successfully with a JSON parser, so they ship straight to a database write with no further checks. Weeks later, a bug report shows a saved record where the phone-number field parsed fine as a JSON string but actually contained an unrelated value pulled from elsewhere in the input. What is the correct assessment?
V7 · Q26Indirect prompt injection: handling untrusted tool-fetched content
An app lets an authenticated, trusted user ask Claude to fetch a webpage via a fetch_webpage tool, then have Claude summarize it and optionally draft a follow-up email based on what it read. A developer is deciding how to pass the fetched page content into the conversation. Which approach best follows the documented mitigations for indirect prompt injection?
V7 · Q27Prompted instructions are not a security control
An internal tool gives Claude Code shell access to run build commands for developers. To prevent destructive actions, the team adds one line to the system prompt: 'Never run rm -rf, never delete the .git directory, never force-push.' They consider this sufficient and make no other changes. Later, an adversarial test prompt gets Claude to run a destructive command anyway. What is the correct assessment?
V4 · Q28Debugging: isolating which layer is actually failing
A support-bot agent has a lookup_order(order_id) tool that queries an internal order database. A user reports that when they ask about the status of order 48291, the bot confidently returns the status of a completely different, seemingly random order. The team's first instinct is to rewrite the system prompt with a stronger instruction like 'always double-check the order ID you were given.' What should they verify before doing that, and why is jumping straight to a prompt rewrite risky?

All mock exams Back to study guides