How to Use Generated Knowledge Prompting

Generated knowledge prompting splits a question into two passes. In the first pass, you ask the model to produce a handful of true, relevant statements about the topic — the background knowledge a careful person would recall before answering. In the second pass, you feed those statements back and ask the model to answer the original question while explicitly using them.

The intuition from Liu et al. 2021 is that language models hold a lot of latent knowledge, but a single-shot answer doesn't always activate the relevant pieces. Asking the model to write the knowledge down first makes it available as context, so the final answer is conditioned on stated facts rather than a snap judgment.

Unlike retrieval-augmented generation, the knowledge here is generated by the model, not fetched from a database. That makes it fast and dependency-free, but it also means the generated facts can be wrong — so the technique works best where the model's underlying knowledge is solid and the failure mode is forgetting to use it, not lacking it.

It earns its keep on questions where the answer depends on background facts the model knows but might skip:

**Commonsense reasoning.** The original target in Liu et al. 2021 — questions like "Can you fit a giraffe in a refrigerator?" where the right answer hinges on facts about size and physical constraints.

**Domain explanations.** "Why does X cause Y?" answers improve when the model first lists the relevant mechanisms, then reasons from them.

**Comparisons and trade-offs.** Asking the model to enumerate the properties of each option first, then compare, produces more balanced output than a one-shot verdict.

**Drafting where omissions are costly.** Briefs, checklists, and explainers benefit when the model surfaces the considerations before writing — it is less likely to leave out something important.

These three techniques are easy to confuse because all of them add context before the answer. The difference is where the context comes from and what it contains.

**Chain-of-thought** generates reasoning steps — the process of getting to the answer. **Generated knowledge** generates facts — the inputs the answer should be built from. **RAG** retrieves facts from an external source instead of generating them.

In practice they compose. You can generate knowledge, then reason over it step by step. And when accuracy matters and the facts are external or volatile, prefer retrieval — see how to combine RAG and prompts for that pattern. On modern reasoning models (GPT-5.5 thinking mode, Claude extended thinking), some of this happens internally, so treat explicit generated-knowledge prompting as most valuable on faster, non-reasoning tiers.

Here is a single-shot prompt that often produces a confident but shallow answer:

``` Is it safe to store cooked rice in the fridge and reheat it the next day? ```

Now the generated knowledge version, run as two passes (or as one prompt with both instructions):

``` Step 1 — Generate knowledge: List 4-5 factual statements about cooked rice, bacterial growth (e.g., Bacillus cereus), safe cooling, refrigeration time limits, and reheating. Step 2 — Answer: Using only the statements above, answer: Is it safe to store cooked rice in the fridge and reheat it the next day? Note any conditions that must be met. ```

The two-pass version forces the model to surface the cooling-time and time-limit facts before answering, so the response includes the conditions a one-shot answer often omits. For food-safety or health questions, always verify against an authoritative source — see the disclaimer below.

This guide is informational and not medical, legal, or food-safety advice. Generated knowledge prompting can surface plausible-sounding statements that are incorrect, because the facts are produced by the model rather than retrieved from a verified source. Never input personal health information, client-confidential data, or other sensitive PII into a chatbot, and verify any high-stakes output with a licensed professional or an authoritative reference before acting on it.