Generated Knowledge & Prompt Priming
Have the model write down relevant facts before answering — surfacing latent knowledge into context, where it constrains the answer.
Generated-knowledge prompting splits a question into two beats: first ask the model to write down facts relevant to the question — no answering yet — then ask the question with those facts sitting in context. It sounds like a parlor trick; the mechanism is anything but. Knowledge a model 'has' stays latent until tokens make it explicit, and an answer generated cold gets no benefit from facts the model never surfaced. Once the relevant facts are in the visible context, the answer must condition on them — considerations a one-shot answer would have skipped are now physically in front of the model as it generates.
You can run it as one prompt ('First, list the facts relevant to this question. Then, using those facts, answer.') or as two calls. The two-call version has a superpower: you can inspect and edit the knowledge between calls — delete the wrong fact, add the missed constraint, or hand the vetted list to a cheaper model for the final answer. This is also where the technique quietly becomes architecture: swap 'model generates facts' for 'retrieval fetches documents' and you've reinvented RAG. Generated knowledge draws on the model's memory and can be wrong; retrieval draws on your sources and can be stale. Production systems often want both.
The adjacent habit is priming: front-loading context before the ask, because the model cannot use what it hasn't seen. A glossary of your team's terms, the audience, the house style, the three decisions already made — pasted before the question, these cost pennies and quietly prevent whole classes of wrong answers. Most 'the model doesn't get it' complaints are priming failures: the asker held the context in their head and never typed it. A useful drill before any complex request: ask what a smart new hire would need to be told first — then tell the model exactly that.
The failure mode deserves respect: confidently generated wrong 'knowledge' is worse than none, because it now sits in context wearing the costume of established fact, and the answer will faithfully build on it. Hallucinated premise in, polished garbage out. So calibrate by stakes. For brainstorming and low-stakes analysis, generate freely. For anything factual that matters, verify the generated facts before the second call — or replace generation with retrieval from a source of truth. Module 3 goes deep on why models fabricate and how to catch it before it compounds.