Instructions and covered skills
Read each prompt-design scenario carefully before selecting a response. Focus on what the model should do with retrieved material rather than relying on outside knowledge. Keep track of constraints such as source boundaries, citation requirements, and output formats. Avoid distractions and turn off notifications before beginning. Choose the option that best supports reliable, grounded responses. Review wording closely when options differ by a small but meaningful instruction.
Key Areas
This test covers prompt patterns for retrieval-augmented generation, where a language model receives relevant documents alongside a user request. Candidates work with instructions that establish the role of retrieved context, limit answers to supported evidence, and prevent the model from filling gaps with unverified information. Key areas include context delimiting, source prioritization, citation placement, handling contradictory passages, and responding when the retrieved material does not answer the question.
Strong performance also requires attention to instruction hierarchy. A well-designed prompt distinguishes application rules, user requests, and retrieved documents so that text inside a source cannot silently change the task. Candidates should recognize how to request structured outputs, such as an answer followed by cited evidence, without encouraging fabricated citations. They should also understand when a concise synthesis is appropriate and when the model should preserve uncertainty, quote exact language, or ask for clarification.
Recommended Preparation
Practice reviewing document-grounded prompts for ambiguous language. Rewrite instructions so they state what sources may be used, how claims should be supported, and what the model should say when support is absent. Work with sample knowledge-base articles, policies, product documentation, and research excerpts that contain overlapping or inconsistent information. Compare prompts that ask for a general answer with prompts that require claim-level citations and explicit uncertainty.
Prepare by designing output templates for common workflows, including support responses, policy summaries, document comparison, and question answering. Pay particular attention to source labels, delimiters, relevance rules, and citation conventions. Consider failure cases such as irrelevant retrieval, contradictory documents, embedded instructions inside a document, and requests that cannot be answered from the supplied context. The goal is to create prompts that produce useful answers while making the model's evidence boundaries clear to the reader.