Claude Models, Prompting, and Context Engineering
Free Claude Certified Architect - Professional practice — 6 questions on Claude Models, Prompting, and Context Engineering, with explanations. No sign-up. Full 12-question mixed test →
A company processes 2 million customer support tickets per day. 99% require simple intent classification into one of 10 categories with sub-second latency and minimal per-ticket cost. The remaining 1% are escalations requiring multi-step reasoning across policy documents. Which architecture BEST fits this scale?
At 2M requests/day, cost and latency dominate for the 99% simple case — Haiku is purpose-built for high-throughput, low-latency classification. Tiered routing reserves the more expensive, higher-latency model only for the small subset of genuinely complex tasks, which is the standard production pattern for model selection at scale.
You are implementing extended thinking via the Messages API for a complex multi-step agentic task that also issues tool calls. Which statement about the required configuration is accurate?
Extended thinking requires a budget_tokens value of at least 1024, and while it is enabled the API does not allow overriding temperature/top_p/top_k (sampling is fixed), since those parameters would undermine the deterministic reasoning behavior the thinking budget relies on.
An agentic coding assistant runs long tool-use loops (file reads, searches, edits) that accumulate large tool results, risking context window overflow after roughly 50 turns mid-task. Which approach best manages this at production scale?
Context editing automatically clears stale tool_use/tool_result content once a threshold is reached, keeping the active context within budget, while the memory tool lets the agent persist important state (progress, decisions, file state) outside the context window so it survives clearing or resets — together this is the standard pattern for long-running agentic loops.
A production pipeline sends thousands of short requests per hour that all share an identical 50K-token system prompt and tool definition block, with only a small user message varying per call. Which technique BEST reduces cost and latency for this specific pattern?
Prompt caching is purpose-built for exactly this pattern: large, identical, reused context (system prompt + tool definitions) that is static across many calls. A cache_control breakpoint after that static block lets subsequent calls reuse the cached prefix at significantly reduced cost and latency rather than reprocessing 50K tokens every time.
An automated pipeline requires Claude's output to strictly conform to a specific JSON schema with zero freeform preamble text, feeding directly into a downstream parser. Which configuration correctly enforces this in production?
Forcing tool_choice to a specific tool whose input_schema mirrors the desired output guarantees the model emits a structured tool_use block matching that schema, with no freeform preamble — this is the reliable, API-enforced mechanism for structured output, rather than relying on the model's adherence to natural-language instructions.
During a multi-step agentic tool-use turn, what does 'interleaved thinking' refer to in Claude's extended thinking feature?
Interleaved thinking lets Claude emit thinking blocks in between sequential tool calls within a single turn, so it can reason about the result of one tool call before deciding whether/how to invoke the next — this capability must be explicitly enabled via a beta header, distinguishing it from standard extended thinking, which reasons once before the full response.
Ready for the real thing?
The full course has two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed answer explanations.
Start my full course on Udemy →