Generated Knowledge Prompting

Generated knowledge prompting is a two-step pattern where you first ask the model to produce relevant facts or background about the question, then feed that self-generated knowledge back in as context before asking for the final answer. Introduced by Liu et al. (2022), it was designed for tasks needing commonsense or world knowledge the model "knows" but doesn't reliably surface in a single pass. Making the knowledge explicit first gives the model a firmer footing for the actual reasoning. For SaaS builders it's a lightweight alternative to retrieval when you have no external corpus to search: instead of RAG, you elicit the model's own latent knowledge as an intermediate step. A worked shape: "List 3-5 key facts relevant to answering X" → then "Using these facts, answer X." The obvious risk is that generated knowledge can itself be wrong — you're not grounding against a trusted source — so it complements rather than replaces retrieval when factual accuracy is critical, and it works best on reasoning where plausible context helps more than verified citations.