Practice · 22 questions

CCDV-F — Mock 3 · Comprehensive

A third independent full attempt — Batches API ceilings, model-tier selection under a capability requirement, MCP primitive selection, headless CI usage, least-privilege tool scoping, and regression-testing discipline for prompt changes. 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 22 answered

V2 · Q1Message Batches API output ceiling vs synchronous calls
A data pipeline team needs to generate long-form summaries (each response frequently exceeding 128K output tokens) for 40,000 documents overnight, with no need for immediate results. Which approach fits the constraints?
V2 · Q2Choosing among official SDK languages
A team is building a backend service in Go that will call the Claude API directly and wants streaming, typed errors, and built-in retry behavior out of the box, without writing custom HTTP handling. What's the right approach?
V2 · Q3Code review: reinventing the agentic tool-call loop
While reviewing a pull request for a new Claude-powered feature, you find that the author wrote a custom while-loop that manually inspects each API response for tool-use requests, executes the corresponding function, appends the result, and repeats — for a small set of simple custom tools with no unusual control-flow needs. What's the most appropriate review feedback?
V2 · Q4Pinning inference geography with a request parameter
A regulated customer requires that all model inference for their workload run within a specific geography for data-residency reasons. Which mechanism addresses this at the request level?
V2 · Q5Consecutive same-role messages are auto-combined
A support-bot harness appends three separate user-role messages to the messages array in quick succession — an initial message followed by two rapid follow-up clarifications — before making a single API call, without merging them into one message object first. What happens on that request?
V2 · Q6Reusing a code-execution container across requests
An application uses the code execution tool across multiple sequential requests in the same session — first to install a package and create a data file, then in a later request to read that file and produce a table. Between requests, what should the application do to preserve the installed package and file across the calls?
V5 · Q7Fable 5 vs Opus 5 for a genuinely highest-capability task
An engineering team is scoping a one-off, extremely difficult analysis where getting the answer right matters far more than cost or latency, and it genuinely requires the single highest level of model capability available. Which model choice fits, and why?
V5 · Q8The output_config.effort parameter replacing budget_tokens
A developer migrating a workload from a fixed-token thinking-budget approach to Fable 5, Opus 5, or Sonnet 5 wants to control how much a model thinks without specifying an exact token count. What is the correct mechanism on these models?
V5 · Q9Cost trade-off: Sonnet 5's price drop vs Sonnet 4.6
A team currently running a high-volume production workload on the prior-generation Sonnet 4.6 ($3/$15 per MTok) is evaluating a move to Sonnet 5, priced at $2/$10 per MTok — a real reduction versus Sonnet 4.6 — while still targeting the best balance of speed and intelligence in its tier. What's the most accurate takeaway for this migration decision?
V1 · Q10Agent SDK language support and CLI fallback
A team at a logistics company wants to build a custom fulfillment-tracking agent in Go, reusing the same permission system, tool-use loop, and session-resume behavior that Claude Code provides. The Claude Agent SDK officially ships packages for Python and TypeScript only. What is the correct way for the Go team to get equivalent agent behavior?
V1 · Q11Orchestrator-workers vs routing for unpredictable decomposition
An engineering team is building an agent to migrate a legacy codebase to a new framework. The number of files needing changes, and the specific nature of each change, can't be known until the agent inspects the repository - some files need simple renames, others need substantial rewrites, and a few need no changes at all. Which workflow pattern best fits, and why?
V1 · Q12External memory for long-running agent sessions
An agent is running a multi-hour task: auditing a large monorepo for a specific class of security issue, file by file. Partway through, its context window fills with details of files already reviewed, risking that it forgets earlier findings or repeats work already done. What is the most effective mitigation?
V8 · Q13Choosing between MCP Tools, Resources, and Prompts
A company builds an MCP server giving Claude access to their internal engineering wiki. Part of the wiki is a static, rarely-changing style guide that should be available to Claude as background context when relevant, surfaced by the application rather than fetched only when the model decides to call something. Which MCP primitive best fits exposing the style guide?
V8 · Q14strict schema conformance for tool arguments
A custom tool for creating calendar events keeps causing runtime crashes: the model occasionally returns an input object missing a required field or with a field of the wrong type, and the downstream handler throws before it can even validate anything. Which fix most directly addresses the root cause?
V3 · Q15Headless/CI usage of Claude Code
A team wants to run Claude Code as a CI step that auto-generates a changelog entry from a diff on every merge to main. The run needs to be fast, deterministic, non-interactive, and free of unrelated hooks, skills, or MCP servers being auto-discovered from whatever happens to be in the repo. Which combination best fits this use case?
V6 · Q16Requesting structured output via output_config
A team is building a pipeline that extracts contact information from support tickets into a database. They need every response to be valid JSON matching an exact schema (name, email, plan tier). Which approach reflects the documented mechanism for this, and how should it relate to the app's own validation logic?
V6 · Q17Context-window budget planning in a multi-turn conversation
A team is building a long-running multi-turn support assistant. Early in development they simply appended every new turn to the growing message history and resent the whole thing on each request, with no plan for what happens as the conversation approaches the context limit. Which practice reflects sound context-window budget planning for this scenario?
V6 · Q18Explicit decision criteria for a classification task
A support-ticket triage system's prompt tells Claude to "flag urgent tickets appropriately." In testing, similar tickets get inconsistent urgency labels from one run to the next. Which revision most directly addresses the cause?
V7 · Q19Least-privilege tool scoping
An internal agent is built to draft and post scheduled Slack status updates. The engineer gives it a general-purpose bash tool with unrestricted shell access, reasoning that bash can do anything the agent might eventually need. Which critique best reflects least-privilege tool design?
V7 · Q20App Attest token expiry as a security property
A mobile app uses App Attest to authenticate its calls to the Messages API from an iOS client, instead of embedding a static API key in the app binary. A developer new to the project asks why App Attest tokens expire after about an hour instead of being issued with a longer lifetime for convenience. What is the best answer?
V4 · Q21Regression testing discipline for a prompt change
A team edits their coding assistant's system prompt to fix a bug where Claude sometimes skipped running tests before declaring a task complete. The fix works well against the specific scenario they wrote to reproduce the bug. Two weeks later, users report Claude is now over-verifying — rerunning tests repeatedly even on tasks where that wastes time. What does this outcome most directly point to?
V4 · Q22Structured logging for after-the-fact layer isolation
When a support agent's response is wrong, an engineer currently can only see the final text Claude returned — no record of which tools ran, what they returned, or why the turn ended. Which change would most improve the team's ability to diagnose future failures by layer (API, tool execution, prompt/context, or model output)?

All mock exams Back to study guides