Generated Knowledge Prompting
Tier: Advanced · Task types: Factual accuracy, reasoning priming, expert-level responses · Difficulty: ⭐⭐⭐☆☆
Table of contents
Definition
Generated knowledge prompting asks the model to first generate relevant facts or context about a topic before answering a question about it. By activating domain knowledge explicitly before the task, the model is better primed to produce accurate, specific answers — rather than drawing on whatever associations are most statistically prominent in training.
The technique is a prompt-level analogue to retrieval-augmented generation, using the model’s own knowledge as the retrieval source.
Use Cases
- Accuracy-sensitive factual tasks where the model is prone to confident hallucination in domain areas
- Nuanced opinion questions that benefit from the model first articulating relevant considerations before taking a position
- Writing tasks requiring specific expertise — priming with domain knowledge before writing produces more authoritative output
Weak Prompt
Should a startup raise venture capital or bootstrap?
Weak Output
Whether to raise venture capital or bootstrap depends on your goals, runway, and market opportunity. VC funding provides capital and network but dilutes ownership…
Assessment: Generic advice that provides no new insight. The model defaulted to surface-level considerations without engaging with the structural trade-offs.
Improved Prompt (Two-Step Generated Knowledge)
Step 1 — Generate Knowledge:
List 8 key facts about the structural differences between venture-backed and
bootstrapped startup trajectories, including typical growth rates, dilution,
time-to-exit, founder control, and failure modes. Be specific and cite
realistic ranges where possible.
Step 1 Output (Gemini):
- VC-backed companies typically raise at 20–35% dilution per round; founders often own 10–20% at Series B.
- Bootstrapped SaaS companies median ARR growth: ~50% YoY. VC-backed: pressure for 3× annual growth.
- Median time to VC exit: 8–10 years; bootstrapped founders can take profits from year 1.
- VC funds require a “venture-scale” outcome (10×+ return) — most markets don’t support this.
- Bootstrapped companies have 90%+ founder decision autonomy; VC-backed cede board seats typically from Series A.
- 75% of VC-backed startups fail to return investor capital; many bootstrapped businesses are cash-flow positive.
- VC is structurally suited to winner-take-all markets; bootstrapping suits niche, sustainable, or services businesses.
- Down rounds and bridge notes are common VC dynamics that create significant founder dilution beyond initial terms.
Step 2 — Answer the Question:
Using these facts as your foundation, give a structured recommendation on
when a startup should raise VC vs. bootstrap. Be specific and decisive —
don't hedge everything.
Step 2 Output (Gemini):
Bootstrap if: You’re building in a niche, capital-efficient, or services market; you value founder control; your business can reach profitability within 24 months on revenue alone; or you want optionality over exit timing.
Raise VC if: You’re competing in a winner-take-all market where speed of scale determines who survives; your business requires significant upfront capital that revenue cannot fund; and you accept that you’re optimising for a large-exit scenario — not lifestyle income.
The honest default: Most startups should not raise VC. VC is a product with specific terms (high growth, large exit, dilution) that suits a narrow set of businesses. The bootstrapped path is harder in year 1 and more controlled in year 5.
Assessment: The answer is specific, decisive, and grounded in real structural considerations activated in Step 1.
Verdict
Generated knowledge dramatically improves the specificity and accuracy of responses in domains where the model holds relevant knowledge but doesn’t automatically surface it in a single-step prompt. The technique is essentially asking the model to do its own research before answering — the overhead is one extra prompt, and the output quality gain is significant.
Common Pitfalls
1. Hallucinated “facts” in Step 1 The model may generate plausible but inaccurate knowledge. Verify key figures independently, especially for quantitative claims.
2. Not referencing Step 1 in Step 2 The second prompt must explicitly instruct the model to use the generated knowledge as a foundation — otherwise it may ignore it.
3. Using generated knowledge on simple tasks Adding a knowledge-generation step to a simple factual question adds latency and cost without benefit. Reserve for complex, nuanced, or domain-specific tasks.