CertForge

1 / 10 single answer medium

An aviation-MRO repository needs a subagent to compare maintenance-manual revisions and report changed inspection intervals. It should not alter source files, execute shell commands, or access unrelated customer records. How should the subagent be defined?

A. Let it update the manuals directly so the parent agent receives cleaner inputs. B. Give it only the manual-search and read tools, constrain it to approved document paths, and require a structured change report. C. Give it the parent agent's full tool set but state in its prompt that it should remain read-only. D. Allow Bash access because the subagent may discover a useful parsing script on its own.

2 / 10 single answer medium

When providing a very long document to Claude and asking for a summary, why is wrapping the document in `<document>...</document>` tags considered a best practice?

A. It forces the model to use XML in its response. B. It reduces the number of tokens required for the prompt. C. It helps the model distinguish between instructions and the provided data. D. It makes the prompt more readable for humans.

3 / 10 single answer hard

Production metrics show that when resolving complex billing disputes or multi-order returns, customer satisfaction scores are 15% lower than for simple cases—even when the resolution is technically correct. Root-cause analysis shows the agent provides accurate solutions but inconsistently explains rationale: sometimes omitting relevant policy details, sometimes missing timeline info or next steps. The specific context gaps vary case by case. You want to improve solution quality without adding human oversight. What approach is most effective?

A. Add a self-critique stage where the agent evaluates a draft response for completeness—ensuring it resolves the customer’s issue, includes relevant context, and anticipates follow-up questions. B. Implement few-shot examples in the system prompt showing complete explanations for five common complex case types, demonstrating how to include policy context, timelines, and next steps. C. Add a confirmation stage where the agent asks “Does this fully resolve your issue?” before closing, allowing customers to request additional information if needed. D. Upgrade the model from Haiku to Sonnet for complex cases, routing based on a defined complexity metric.

4 / 10 single answer medium

In an orchestrator-workers deployment for competitive analysis, three parallel workers research the same vendor from different source sets and return conflicting claims about its pricing model. The orchestrator currently concatenates all three reports. Downstream consumers complain the output contradicts itself. What should the orchestrator do instead?

A. Keep the report of whichever worker used the most sources and discard the others. B. Add an explicit synthesis step: reconcile claims, prefer better-sourced ones, and surface unresolved conflicts as flagged uncertainty rather than silent contradiction. C. Serialize the workers so each sees the previous one's findings and conflicts never arise. D. Constrain all workers to one shared source list so their outputs can't conflict.

5 / 10 single answer hard

A telecom assistant must reserve a mobile number for a customer in a specific rate centre for up to 30 minutes. Which tool definition is most reliable?

A. Name it `activate_number` but mention in the description that activation happens later. B. Name it `reserve_mobile_number`; require structured `rate_centre_code`, `customer_id`, and integer `hold_minutes` constrained from 1 to 30; describe that it creates a temporary hold but does not activate service. C. Name it `number_task` and accept one free-form `request` string. D. Name it `reserve_mobile_number` and allow any object so future fields can be added without schema changes.

6 / 10 select ALL that apply hard

A multiplayer game generates a short in-character paraphrase of an already approved NPC line during live play. Requests are high-volume, the persona instructions are stable, outputs are capped at twelve words, and p95 latency is too high. Select all that apply.

A. Include the full match transcript even when the approved source line is self-contained. B. Ask for a paragraph of hidden backstory with every twelve-word line. C. Assume output token cost is irrelevant because the output is short. D. Evaluate a smaller suitable model on a fixed quality and safety test set before retaining the largest model. E. Measure input processing, output generation, cache performance, and network or tool overhead separately. F. Keep the stable persona and safety prefix identical and cache it when measured hit rates and pricing justify the cache.

7 / 10 single answer medium

A field-service company uses Claude to turn technician notes into customer visit summaries. Instructions describe the format, but outputs vary for no-access visits, partial repairs, subcontractor follow-up, and warranty exclusions. What is the most targeted improvement?

A. Increase temperature so the model can find more natural wording. B. Add multishot examples for those edge cases using the exact headings, omissions, and wording constraints expected downstream. C. Repeat 'be consistent' several times in the system prompt. D. Remove the structured headings and accept free-form narratives.

8 / 10 single answer medium

When an agent's database query tool fails, it currently returns 'Error: query failed' (the stack trace goes to server logs). Agents respond by retrying the identical query several times, then giving up. What should the tool return instead?

A. The full raw stack trace, so the model has maximum information. B. Nothing — have the tool auto-correct common mistakes like near-miss column names. C. HTTP-style codes (e.g. 400) that the model can look up. D. A structured, model-readable error stating what went wrong and what a valid retry looks like — e.g. 'Unknown column customer_nm in table customers. Available columns: customer_name, customer_id, ...'.

9 / 10 single answer hard

A fleet-insurance product has permanent safety rules, insurer-specific coverage definitions, and a broker's one-time question about a claim. Where should these elements be placed?

A. Put safety rules in system, coverage definitions in retrieved context, and the question in user. B. Put safety rules and coverage definitions in system, then append the broker question as context. C. Put the broker question in system and supply coverage definitions as assistant examples. D. Put all three elements in one user message separated by headings for easier prompt maintenance.

10 / 10 single answer hard

A legal research platform needs Claude to trace the evolution of a specific property rights concept across a 2,000-page historical manuscript. What is the most appropriate architectural strategy for managing this context?

A. Stuff the entire 2,000-page manuscript into the system prompt and enable prompt caching to affordably process the massive token count. B. Provide the manuscript via a simple `read_file` tool so the model can read it all at once during its first reasoning turn. C. Index the manuscript into a vector database for semantic retrieval, or implement an agentic recursive summarization workflow. D. Rely entirely on the model's pre-training data, as standard historical manuscripts are already deeply encoded in the model's weights.

Try Claude Certified Architect – Foundations — no account needed