GKP
Make it state the facts before it uses them — then make it say which ones it is unsure of.
What is GKP?
Generated Knowledge Prompting splits one question into two moves: first make the model write down the facts that bear on the question, then make it answer using only those. Same model, same knowledge, two passes — and on factual questions the second pass is measurably better than answering directly.
A model answering directly commits to a conclusion and assembles support for it on the way past. Forcing the facts out first inverts that: the statements get written before there is an answer to defend, so they are less bent toward one. It is the same instinct as showing your working, applied to knowledge rather than to reasoning. And it gives you something to inspect. When the answer is wrong, a direct reply tells you nothing; a numbered list of the statements it was built on tells you exactly which one was false. That is the practical reason to use it, and it is worth more than the accuracy gain.
Where GKP Came From
Liu et al., ACL 2022
"Generated Knowledge Prompting for Commonsense Reasoning" (Jiacheng Liu, Alisa Liu, Ximing Lu, Sean Welleck, Peter West, Ronan Le Bras, Yejin Choi and Hannaneh Hajishirzi, ACL 2022). The method generates knowledge statements from a language model and then supplies them as additional input when answering the question. It needed no structured knowledge base and no task-specific supervision, and set the state of the art on NumerSense, CommonsenseQA 2.0 and QASC.
The finding underneath it: a model knows more than it uses
That is the genuinely interesting claim. If asking a model to state relevant facts, and then answer, beats asking it to answer — the knowledge was there and the single pass was not reaching it. The paper frames large models as flexible sources of external knowledge, which is a strange and useful way to think about it: you are using the model to retrieve from itself.
The gap the technique leaves open
Generated knowledge is generated, not retrieved. Nothing in the method checks whether a statement is true, and a false statement in stage one is reasoned from as confidently as a true one — it will look identical on the page. Adding a confidence pass is not part of the published technique and it is the single most valuable thing you can bolt on, because it converts an unmarked list of claims into a list with the risky ones labelled.
The 4 Slots, One at a Time
Each slot is a decision. Leave it out and the model still makes it — just without you.
Put the question in, but instruct that no attempt at an answer may appear before the knowledge section. Without that the model answers in the first line and produces the "knowledge" afterwards as justification, which is the failure mode the technique exists to prevent and it is easy to miss because the output looks right.
Ask for a count — eight to twelve is a useful band — and insist on one fact per statement. Prose blurs claims together so that a false one hides inside a true sentence; numbered atomic statements can be checked, quoted and argued with individually. Say you want rules, thresholds, definitions and exceptions rather than general context.
Not in the original technique, and the reason to use this page. Mark each statement, and for anything below HIGH say what would need verifying and where. It costs almost nothing and it turns the output from a wall of equally-confident claims into a short list of things to check. Treat the labels as a relative ranking within one answer, not as probabilities — they are not calibrated.
Require the answer to cite which statements it relied on, and to surface any dependency on a LOW-confidence one in the answer rather than in a footnote nobody reads. That single rule is what stops a shaky premise disappearing into a confident conclusion.
One Task, Before and After
The task: A plain-English explainer on VAT for cross-border digital services. Both prompts below are scored by our free prompt checker — paste either in and you will get the same number, because the scoring is deterministic.
Explain how VAT works for digital services.
Seven words. You get a confident, well-organised explanation containing two thresholds that are out of date and one rule that applies to a different jurisdiction, presented in exactly the same tone as the parts that are correct.
Answer a question about VAT on cross-border digital services, but generate the relevant knowledge first and answer only afterwards. Stage 1 - Knowledge. Before attempting any answer, write eight to twelve numbered statements of the specific rules, thresholds, definitions and exceptions that bear on this question. One fact per statement. Treat these as working notes, not prose. Stage 2 - Confidence. This is the stage the technique needs and does not include: mark each statement HIGH, MEDIUM or LOW confidence, and for anything below HIGH say what would need checking and where. Generated knowledge is generated, not retrieved - a wrong statement here will be reasoned from as though it were true, and it will look exactly like a right one. Stage 3 - Answer. Only now answer the question, citing the numbered statements you relied on. If the answer depends on a statement you marked LOW, say so in the answer rather than in a footnote. The question: a UK company sells a subscription analytics product to business and consumer customers in the EU, the US and Australia. Where is VAT or its equivalent due, who accounts for it, and what changes when the customer is a business rather than a consumer? Reason from the statements before you answer - the knowledge section is working notes that the answer must be built on, not a preamble to it. Produce a plain-English guide of 700 to 900 words for a founder with no finance background who needs to know what to ask an accountant, not to file anything themselves. For example, a knowledge statement should look like this: ``` 7. For B2B sales of digital services into the EU, the reverse charge generally shifts the VAT accounting obligation to the customer. [HIGH - well established, but the customer's VAT number must be validated] ``` Ensure the answer names the specific decision points a founder must resolve, and prioritise the rules that change the answer over background. Do not state a rate or a threshold you are not confident in without marking it LOW, do not present this as advice, and end by naming the three questions to put to an accountant.
Ninety-one. The confidence pass is the addition that matters: it does not make the model more accurate, it makes the inaccuracies findable — which on a tax question is the more useful property.
Ninety-one, and no score can tell a true statement from a false one
One check is out of reach, and the risk the whole technique is aimed at is the one nothing measures:
No persona slot. Adding "you are a VAT specialist" is worth eight points and is arguably counterproductive here — an expert persona raises the confidence of the delivery without raising the accuracy of the content, which is the wrong direction for a technique built around flagging uncertainty.
The whole point, and completely invisible. Twelve fluent, well-numbered, confidently-worded false statements score exactly as well as twelve true ones. GKP makes the claims inspectable; it does not make them right, and no rubric closes that gap.
A model's HIGH is not a probability. It is a useful ranking within one answer — the LOW items really are the ones to check first — and it is not evidence that the HIGH items are safe.
So the instruction worth adding to any factual prompt, GKP or not: make it mark what it is unsure of, and say what would need checking. It earns no points, it does not improve accuracy, and it changes an unverifiable answer into a three-item to-do list for someone who can verify it.
Copy-Paste Prompt Template
Replace the bracketed placeholders with your specific details.
[The question — and an explicit instruction that NO answer may appear before the knowledge section] STAGE 1 — Knowledge: [8-12 NUMBERED statements. One fact each. Rules, thresholds, definitions, exceptions. Working notes, not prose] STAGE 2 — Confidence: [mark each statement HIGH / MEDIUM / LOW, and for anything below HIGH say what would need checking and where] STAGE 3 — Answer: [built from the statements, citing which ones. If it depends on a LOW statement, say so IN the answer, not in a footnote] [Reader and length. End with the questions to put to someone who can verify]
When GKP Fits — and When It Does Not
- Factual questions where being wrong is expensive — tax, regulation, technical rules.
- Explainers you will hand to someone who cannot check the content themselves.
- Any answer that will be verified afterwards, since the statements are what gets checked.
- Domain questions where you want to see the model's premises before its conclusions.
- Cases where you suspect the model knows more than a direct answer is surfacing.
- Anything with real retrieval available. Use ReAct — generated knowledge is not sourced.
- Creative and subjective work, which has no facts to pre-state.
- Simple lookups, where two passes cost time and add nothing.
- Fast-moving facts — prices, current events, versions. The model is recalling, not checking.
- High-stakes decisions taken on the output alone. This flags uncertainty; it does not resolve it.
10 Ready-Made GKP Prompts
Every prompt below was produced by the Frompting generator with GKP selected — not written by hand for this page. Each is scored by our prompt checker; the median is 88/100. Click one to open it, then copy.
VAT on cross-border digital services 94
You are a tax specialist with expertise in value‑added tax (VAT) for digital services. First, generate a concise list of the essential facts and rules that govern VAT on cross‑border digital services, covering: - Definition of “digital services” for VAT purposes - Determination of the place of supply (seller‑ vs. buyer‑location rules) - Registration obligations for non‑resident sellers - Thresholds that trigger VAT registration in the buyer’s country - Invoicing and reporting requirements (e.g., OSS, MOSS) - Common exemptions or reduced rates that may apply Present these facts as bullet points, citing the relevant EU or OECD principles where appropriate. If any item depends on a specific jurisdiction, note the placeholder needed. Next, using the generated facts, explain how VAT applies to a seller of digital services selling to customers in other countries. Structure the explanation as follows: 1. Identify the seller’s country – **[SELLER_COUNTRY]**: supply the name of the country where the provider is established. 2. Identify the buyer’s country – **[BUYER_COUNTRY]**: supply the name of the customer’s country. 3. Specify the type of digital service – **[SERVICE_TYPE]**: e.g., streaming, software download, SaaS. 4. State any applicable registration threshold – **[VAT_THRESHOLD]** (if known). Provide a step‑by‑step guide (≈ 250–300 words) that tells the seller what actions to take: determine place of supply, assess registration need, choose the appropriate OSS/MOSS scheme, issue compliant invoices, and file returns. Quality criteria: - Factually accurate according to current EU/UK/OECD VAT rules - Clear, actionable steps for the seller - Concise language without unnecessary jargon Assume any missing details are unknown; explicitly state those assumptions and ask up to three clarifying questions (e.g., “What is the seller’s country of establishment?”) before finalizing the guidance. Exclude discussion of other taxes (e.g., income tax, customs duties) and focus solely on VAT for digital services. Write this for [AUDIENCE: who will read the output, and how much they already know]. Match the depth, vocabulary and examples to that reader.
ISAs versus a general investment account 77
You are a financial education specialist. Your task is to explain the difference between an ISA (Individual Savings Account) and a general investment account for a reader seeking clear, accurate information. First, list the key factual elements that define each account type, including: - tax treatment (e.g., tax‑free growth, tax‑deferred vs. taxable earnings) - contribution limits or caps, if any - eligible investment vehicles (stocks, bonds, funds, etc.) - withdrawal rules and penalties - typical use cases and regulatory environment. Second, synthesize those facts into concise knowledge statements that compare the two accounts across the dimensions above, highlighting the practical implications for an investor’s portfolio planning. Finally, compose a cohesive explanation that: - presents the comparison in plain language without jargon, defining any necessary technical terms on first use; - uses bullet points for the side‑by‑side comparison, followed by a short narrative summary (≈150–200 words total). Quality criteria: factual accuracy, clarity for a non‑expert audience, and logical flow. Exclude personal financial advice or recommendations; focus solely on describing the structural differences.
How DNS propagation actually works 84
You are an expert network engineer tasked with clearly explaining how DNS propagation works. First, generate a concise set of factual statements that cover the essential mechanisms involved in DNS propagation, including: - The role of authoritative name servers and recursive resolvers - How DNS records are cached and the impact of TTL values - The sequence of events from a DNS record update to its visibility across the internet - Common factors that cause propagation delays Present these facts as a short bullet-point list (no more than 8 items), ensuring each point is technically accurate and uses correct terminology. Next, using the generated facts as the foundation, craft a clear, step-by-step explanation of DNS propagation aimed at readers with a basic understanding of internet concepts. Structure the final answer with: 1. A brief introductory sentence (1-2 lines) that sets the context. 2. A numbered list (4-6 steps) that walks the reader through the propagation process, referencing the facts you generated. 3. A concluding sentence that summarizes why propagation time can vary. The entire explanation should be between 305-405 words, written in plain language while retaining technical precision. Ensure that the content stays focused on DNS propagation and does not diverge into unrelated DNS topics.
GDPR subject access requests 92
You are a knowledgeable privacy‑law specialist. Your audience is a [AUDIENCE: specify the reader’s role or expertise level, e.g., “junior compliance officer” or “senior data‑protection officer”]. First, generate a concise list of the essential factual elements that define a GDPR Data Subject Access Request (DSAR), including the legal basis, time limits, verification steps, and typical data categories involved. Next, using those generated facts as grounding, produce a clear, step‑by‑step explanation of how to handle a DSAR from start to finish. Structure the answer with the following sections: 1. **Overview** – brief definition and purpose (≈30 words). 2. **Key Legal Requirements** – bullet points of the core obligations (≤5 bullets, each ≤20 words). 3. **Process Flow** – numbered steps covering receipt, identity verification, data collection, review, and delivery, with any mandatory timeframes noted. 4. **Practical Tips** – concise recommendations to avoid common pitfalls (≤4 bullets). The entire response should be between 250 and 350 words. Quality criteria: - All statements must be grounded in the GDPR text or widely accepted guidance; mark any approximate figures as “approx.”. - Use plain language appropriate for the specified audience, defining any legal terms on first use. - Present the process in a logical order that a practitioner could follow directly. Exclude any discussion of unrelated regulations (e.g., HIPAA, SOC 2) and do not provide sample legal clauses or templates. If any of the required details are unclear, state your assumptions explicitly and ask up to three clarifying questions before delivering the final explanation.
How mortgage interest is calculated 82
You are a financial education specialist. Your task is to produce a clear, step‑by‑step explanation of how mortgage interest is calculated, suitable for a general audience seeking to understand the basics of mortgage financing. First, generate a concise list of the essential facts and concepts needed to explain mortgage interest calculation, including: - the definition of mortgage interest and how it differs from principal - common interest rate types (fixed, variable, APR) and how they are expressed - the standard formula(s) used (e.g., simple interest, amortizing loan formula) and the meaning of each variable (principal, rate, term, payment frequency) - typical assumptions about compounding periods and payment schedules Next, using only the facts you have just generated, craft the final explanation. Structure the answer as follows: 1. **Overview** – one short paragraph summarizing what mortgage interest is. 2. **Key Variables** – a bullet list defining each variable that appears in the calculation. 3. **Calculation Steps** – a numbered sequence that walks the reader through the computation, showing how to plug values into the formula and how payments are allocated between interest and principal. 4. **Example** – a concrete numeric example using a [LOAN_AMOUNT: specify the principal amount] loan, a [INTEREST_RATE: specify the annual interest rate] rate, a [TERM_YEARS: specify the loan term in years], and a [PAYMENT_FREQUENCY: specify monthly or other] payment schedule. Show the first month’s interest amount and the resulting principal reduction. The entire response should be between 250 and 350 words. Quality criteria: - Accuracy: all formulas and definitions must be correct. - Clarity: use plain language, avoid jargon unless defined, and keep sentences short and varied. - Practicality: the example must be realistic and illustrate the calculation clearly. Do not include any content beyond the scope of mortgage interest calculation (e.g., tax implications, insurance, or refinancing). If any required detail is missing, state your assumption explicitly before proceeding.
UK employment statuses compared 77
You are a knowledgeable UK employment‑law specialist. First, generate a concise, factual list of the main UK employment statuses (e.g., employee, worker, self‑employed, contractor, agency worker, zero‑hours contract holder), including for each status: the key legal rights and obligations, typical contract features, and any statutory tests that determine the classification. Present this information as a markdown table with columns “Status”, “Core Rights”, “Typical Features”, and “Legal Test”. Next, using only the facts you have just generated, write a clear explanation that compares these statuses, highlighting the most important differences in terms of tax treatment, employment protections, and eligibility for benefits. Structure the comparison as a short narrative of 250–300 words aimed at [AUDIENCE: specify the intended reader, e.g., “HR professionals”, “small‑business owners”, or “general public”]. **Quality criteria** - All statements must be accurate according to current UK law (cite the relevant act or regulation where appropriate). - The explanation should be easy to understand for the specified audience, using plain language while retaining correct legal terminology. - The output must not provide personalized legal advice; include a brief disclaimer to that effect. If any required detail is unclear, state your assumption explicitly and ask up to three clarifying questions before proceeding. Use this table shape, one row per item, filling values from your analysis: | “Status” | “Core Rights” | “Typical Features” | And “Legal Test” | | --- | --- | --- | --- |
What happens in a card transaction 92
You are a knowledgeable financial technology educator. Your task is to produce a clear, step‑by‑step explanation of what happens during a payment card transaction, written for a general audience with no prior technical background. First, generate a concise list of the essential facts and technical components involved in a typical card‑present transaction (e.g., cardholder, merchant, acquiring bank, issuing bank, payment network, authorization, settlement). Include brief definitions for each component and note the typical sequence of messages exchanged. Next, using only the facts you just generated, craft a coherent narrative that walks the reader through the entire process from the moment the card is presented to the final settlement, highlighting the role of each participant and the purpose of each step. Output should be approximately 300 words, presented as a flowing paragraph with short bullet points only where they aid clarity (e.g., listing participants). Ensure the explanation is factually accurate, logically ordered, and easy to understand for non‑experts. Quality criteria: 1. Accuracy – all described steps must reflect real‑world card transaction flows. 2. Clarity – use plain language, define any necessary technical terms, and avoid unnecessary jargon. 3. Cohesion – the narrative should smoothly connect each step to the next without gaps. Exclude any speculation about future payment technologies or unrelated payment methods.
How solar feed-in tariffs work 74
You are an expert energy policy analyst. Your task is to produce a clear, concise explanation of how solar feed‑in tariffs (FITs) operate, suitable for a general audience seeking to understand the mechanism, typical policy goals, and common implementation features. First, generate a list of the essential factual elements about solar FITs, including: - the basic definition and purpose of a FIT; - how the tariff rate is set and adjusted; - the typical contract terms (duration, eligibility, metering); - the financial flow from utility to solar producer; - common variations across jurisdictions (e.g., fixed vs. degressive rates, caps, auction mechanisms). Next, using those generated facts, craft the final explanation that: - integrates the factual elements into a coherent narrative; - highlights the key steps a solar system owner experiences from installation to receiving payments; - notes at least two typical policy objectives (e.g., renewable adoption, price stability). The final answer should be 250–300 words, written in plain language with short paragraphs and bullet points where helpful. Ensure the explanation is accurate, avoids technical jargon unless defined, and does not assume any specific country’s policy. If any critical detail (such as the typical tariff adjustment interval or common contract length) is unclear, state your assumption explicitly and ask up to three clarifying questions before finalizing the response. **Quality criteria:** factual correctness, logical flow, readability for non‑specialists. **Boundary:** do not include detailed legal provisions, case‑specific examples, or numerical rate tables.
Carrying forward pension allowance 88
You are a knowledgeable financial advisor specializing in pension regulations. Explain the rules for carrying forward pension allowance, focusing on the key conditions, limits, and timing that determine when and how unused allowance can be transferred to future tax years. [JURISDICTION: specify the country or tax jurisdiction] [ALLOWANCE_TYPE: specify whether the allowance is a personal pension contribution limit, employer‑sponsored limit, or another category] [CARrying_FORWARD_LIMIT: specify the maximum amount or percentage that may be carried forward, if known] Your response should be about 250 words, written in clear, professional prose. Include: 1. A concise summary of the overall principle. 2. The specific eligibility criteria (e.g., income thresholds, contribution history). 3. The calculation method for the carry‑forward amount. 4. Any deadlines or filing requirements. Quality criteria: factual accuracy, logical flow, and use of plain language with brief definitions for any technical terms. If any of the placeholders above are unknown, state your assumption clearly and ask up to three clarifying questions before finalizing the explanation. Do not include unrelated tax advice, promotional content, or references to external sources. Write this for [AUDIENCE: who will read the output, and how much they already know]. Match the depth, vocabulary and examples to that reader.
How a certificate authority builds trust 92
You are an expert in public key infrastructure who explains complex security concepts clearly. Your task is to produce a concise, accurate explanation of how a certificate authority (CA) establishes trust. **Audience:** [AUDIENCE: specify the intended readers, e.g., “technical professionals”, “students with basic networking knowledge”, or “non‑technical stakeholders”]. **Desired length:** approximately 300 words, formatted as a short essay with clear paragraph breaks. First, generate a list of the essential factual points about CAs, including: the role of root and intermediate CAs, the issuance and signing process, certificate chaining, and how browsers and operating systems use trusted root stores. Next, using those generated facts, craft the final explanation that: 1. Presents the information in a logical sequence that builds understanding step‑by‑step. 2. Uses precise technical terminology while briefly defining any specialized terms for the specified audience. 3. Highlights why each step contributes to establishing trust. **Quality criteria:** - All statements must be verifiable from established PKI standards and widely accepted security literature. - The explanation should be clear, free of unnecessary jargon, and maintain a natural human tone. - Logical flow should enable the reader to follow how trust is built from root certificates to end‑entity certificates. **Boundary:** Do not discuss unrelated topics such as certificate revocation mechanisms, certificate pinning, or detailed cryptographic algorithm internals. If any required detail is unclear, explicitly state your assumptions and ask up to three clarifying questions before completing the explanation.
Scores range from 74 to 94. They are shown as generated rather than cherry-picked — a library where every entry scores in the nineties tells you it was curated, not measured.
GKP vs the Alternatives
The version with real sources. ReAct fetches facts from tools; GKP recalls them from the model. If you have retrieval available, use it — GKP is what you do when you do not.
The sibling technique, aimed at a different failure. CoT makes the reasoning explicit; GKP makes the facts explicit. They combine cleanly on problems that need both.
The natural pairing. GKP surfaces the premises, MCP builds in the self-evaluation stage — together they get you premises plus a critique of them.
For shape rather than substance. Few-shot fixes how an answer looks; GKP addresses what it is built on.
The human completeness checklist. 5W1H asks whether all the facts are present; GKP asks the model to state them before using them.
Five Ways People Get GKP Wrong
The defining GKP failure and easy to miss. Without an explicit instruction, the model answers in the opening line and produces the knowledge afterwards — which is justification, not generated knowledge, and it looks identical on the page.
A paragraph of context blurs claims together so a false one hides inside a true sentence. Numbered, atomic, one fact each — that is what makes it checkable.
Without it every statement carries the same authority, and the one that is wrong is the one that reads exactly like the eleven that are right.
They are not calibrated. Use them to rank what to check first within a single answer, not to decide that the HIGH items need no checking.
If you can look it up, look it up. Generated knowledge is the model recalling — a reasonable fallback and a poor substitute for a source.
Without it a shaky premise disappears into a confident conclusion. Requiring the answer to name the statements it used is what keeps the chain visible.
GKP Questions
What is generated knowledge prompting?
A two-stage technique: have the model write out the facts relevant to a question first, then have it answer using those statements as input, rather than answering directly.
Where does it come from?
"Generated Knowledge Prompting for Commonsense Reasoning" (Liu et al., ACL 2022). It needed no external knowledge base and set the state of the art on several commonsense reasoning benchmarks.
Does it stop the model hallucinating?
No, and it is important to be clear about that. It makes the premises visible so a wrong one can be found, but nothing in the technique checks whether a generated statement is true. That is why the confidence pass is worth adding.
Why does splitting it into two passes help at all?
Because the statements get written before there is a conclusion to defend, so they are less shaped by one. The paper's framing is that a large model is a flexible source of knowledge — you are using it to retrieve from itself.
Should I trust the confidence labels?
As a ranking within one answer, yes — the LOW items are genuinely the ones to check first. As probabilities, no. Model-stated confidence is not calibrated.
When should I use ReAct instead?
Whenever real retrieval is available. GKP recalls from the model; ReAct fetches from sources and cites them. For anything current or high-stakes, a real source beats a recalled one.
Generate a GKP Prompt Instantly
Skip the manual template — Frompting applies GKP to your topic in one click.
Try it FreeFramework Details
| Name | GKP |
| Stands for | Generate-Knowledge-Prompting |
| Domain | AI & Prompt Engineering |
| Steps | 3 |
| Access | Pro |