TL;DR
The method is simple: before the model writes the document, it asks you questions. You answer three clarifying questions (what matters most, what resources you have, who the reader is), pick one of several ready-made strategic options and write a couple of lines on why you chose it. Only then does the model assemble the final text from your answers.
The core pain is the "just ask it" prompt. People write one line, and the model answers the wrong question. In the control group, one in three documents missed the topic because the prompt didn't contain the heart of the problem. Even people who described the problem correctly got smooth, polished texts with no clear position and no honest breakdown of "what we lose, what we gain". The model guesses what you want, and it guesses the average.
First you lock in context, choice and rationale, then generation happens. It costs 2-3 extra minutes. In the experiment, document quality rose by about a third. The biggest gains were in strategic coherence and trade-off analysis. The scheme added no technical specificity (laws, regulations, figures).
Want this for your own case? Apply this study to your task: AI adapts its method to what you are working on and writes the prompt.
Method Blueprint
STEP 1: You describe the task in your own words → a short prompt
STEP 2: The model asks 3 clarifying questions (priority, resources/infrastructure, audience) → you answer
STEP 3: The model offers 3-4 "cards" — strategic position options → you pick one
STEP 4: You write your rationale for the choice in your own words → 2-4 sentences
STEP 5: The model synthesizes the document from everything collected → a finished plan/memo
It can all be done in a single chat. Generation starts only after step 4.
Example in Practice
Task: The owner of a home-goods store on Amazon. The marketplace raised its fees, margins dropped, and there's a meeting with partners in a week. They need a 30-day plan. This is a task about choosing a strategy and making trade-offs, which is where the method shines. If you need precise tax or customs regulations, the method won't help: you need a specialist for that.
Prompt:
I own a home-goods store. 80% of sales go through Amazon. As of the 1st, fees went up and my margin fell from 22% to 11%. I have a meeting with my partners in a week. I need a 30-day action plan.
Don't write the plan right away. Let's work step by step:
1. Ask me exactly 3 clarifying questions: my main operational priority, my current infrastructure and resources, and the main reader of the plan. For each, give 3-4 answer options plus "other".
2. Wait for my answers.
3. Propose 3-4 strategic position options, each with one line of "what we gain / what we lose". Wait for my choice and my rationale.
4. Only after that, write the 30-day plan.
Result: The model will first ask three questions with answer options (for example, what matters more: keeping revenue or restoring margin). Then it will show several strategy "cards", such as "stay and raise prices", "diversify channels", "move to your own Shopify store". Each will have a short "gain / lose". After your choice and rationale, you get a 30-day plan built around your choice. It will include a trade-off breakdown, not a pile of generic advice. Double-check exact figures on fees and taxes yourself.
Why It Works
The model's weakness. It doesn't know your context. If the task is described in one line, the model fills in the gaps with an average guess. The result: either an answer about the wrong thing, or a smooth text with no position. People are also bad at noticing these failures. In the experiment, self-rated quality barely matched the actual scores.
The strength. The model is good at synthesizing material when it's specific and structured. It's also good at asking substantive questions and laying out options with pros and cons. It doesn't need to guess; it needs to assemble.
How the method uses this. You make the decision (the direction and the reasons), and the model writes it up. The document is built around your choice, so it has an explicit position and an honest breakdown of "what we lose". You also stay engaged in the reasoning: participants felt more strongly that they had shaped the plan themselves.
Control knobs: - Number of questions. For simple tasks 1-2 is enough; for complex ones, 4-5. More questions, more time. - Strategy cards. Ask the model to make the options opposite in spirit (aggressive / cautious / middle ground). That makes the choice more meaningful. - Rationale. This is the most "human" part. Skip it and you lose engagement, and possibly some quality too. The paper didn't isolate this, so nobody knows for sure. - Document reader. The audience question sets the tone. The same plan gets written differently for a board of directors and for the team.
Prompt Template
Task: {task and context in 2-3 sentences}
Don't write the final document right away. We'll work in 4 stages.
STAGE 1. CLARIFICATION
Ask me exactly 3 questions:
— Main priority: {what exactly needs clarifying: timeline / margin / risk / reputation}
— Infrastructure and resources: {what I have: people, systems, budget}
— Main reader: {who will read it and make the decision}
For each question, give 3-4 answer options and an "other" option.
Stop and wait for my answers.
STAGE 2. CHOOSING A POSITION
After my answers, propose {number_of_options} strategic position options.
For each: a name, the gist in one line, "what we gain", "what we lose".
Stop and wait for me to pick an option.
STAGE 3. RATIONALE
Ask me to explain in 2-4 sentences why I chose this particular option.
Stop and wait.
STAGE 4. SYNTHESIS
Only now write the {document_type}, based on my answers, the chosen position and my rationale.
Required: explicitly show the trade-offs of the chosen path and how each step connects to the chosen position.
No fluff. Flag separately anything that needs to be checked by a specialist (regulations, figures, legal details).
What to fill in:
- {task}: the situation, with numbers and deadlines.
- You can swap the three questions for your field, but keep the idea: priority → resources → reader.
- {number_of_options}: 3-4.
- {document_type}: a 30-day plan, a memo, a pitch, a strategy.
🚀 Quick start: paste into the chat:
Here is a template for a step-by-step "questions first, then the document" workflow. Adapt it to my task: [your task].
Ask me questions to fill in the fields.
[paste the template above]
The LLM will ask about the task's context, who will read the document and what constraints you have on time and resources. It needs these so that the three clarifying questions and the strategy cards in the template are about your situation, not generic. It will take the pattern from the template and build a working prompt.
Limitations
⚠️ Most of the effect comes from rescuing weak prompts: About 60% of the overall gain is explained by the fact that without the method, documents often missed the topic. If you already describe your task in detail, the gain is more modest. But even for people who described the problem correctly, it was noticeable, about a fifth.
⚠️ Specificity doesn't improve: Technical and regulatory accuracy barely improved. The scheme provides structure and a position, not knowledge. Check laws, figures and standards separately.
⚠️ You won't notice that it got better (or worse): Self-rated quality is barely related to the actual scores, with or without the method. You can't rely on the feeling of "this is a good document".
⚠️ Nobody knows which part works: The study compared the whole package (questions + choice + rationale) against a single prompt. It didn't isolate which element produces the effect.
⚠️ One task, one generation model, a fictional scenario: Everything was tested on a single made-up crisis, and participants were online respondents, not real executives. Applying it to your business is a reasonable hypothesis, not a proven fact.
⚠️ Quality scoring is shaky: The scores from two LLM judges and a human diverged a lot in absolute points and agreed better on the "better/worse" ranking. The experiment wasn't publicly preregistered. The memory effect is a weak signal; it didn't survive robustness checks.
⚠️ The cost is time: On average 2-3 minutes longer. For a simple task, that's overkill.
How They Studied It
Researchers from the Oxford Centre for Impact Research collected 400 sessions on the Prolific platform. Participants were managers and professionals from various countries. Each was given the same fictional crisis: two weeks before a launch in Singapore and Malaysia, the regulator requires data to be stored on local servers, while the company's entire architecture sits in Frankfurt. They had to write a prompt and get an action plan. One half got a response to their prompt right away (196 sessions). The other half went through three questions, a strategy choice and a rationale (204 sessions). The plans were scored on four criteria (coherence, technical specificity, trade-off analysis, density of thought) by two LLMs and one human, none of whom knew which group was which.
The result jumps out right away: all three raters saw an improvement, although the size varied a lot. On average +32% (on a 4-to-40 scale: 15.5 vs. 20.5). The most interesting part is inside the numbers. In the control group, 34% of documents were rated off-topic because the prompt lacked the heart of the problem. In the method group, it was 5%. But when the authors looked separately at participants who described the problem correctly on their own, the method still gave +21%. So the benefit isn't just a "hint about the core issue", but also the structure itself: "chose a position, explained it, got the text". Trade-off analysis and coherence went up, while technical accuracy stayed flat. The method helps you think and structure, but it doesn't replace an expert. Participants in the method group felt more engaged, and it cost them 2.4 minutes.
Adaptations and Extrapolations
🔧 Technique: add a fact-check stage → patch the method's weak spot
The method didn't improve technical accuracy. Add this to the end of the template: "After the document, give a separate list of every claim, figure and regulation I should verify with a specialist or in the primary source, with a note on how confident you are in each". This is my own adaptation; the paper had no such step.
🔧 Technique: drop the ready-made cards → the model builds options from your context
If your task is unusual, in "Stage 2" ask: "Based on my answers, propose 3 position options that differ sharply in risk and speed". That keeps the options from being cookie-cutter.
Extrapolation: moving it into an agent's instructions. The "questions and choice first, then the result" idea can be written into a system prompt or an agent's instruction file. The paper didn't test agents; this is my hypothesis:
For any task involving strategy, planning or architecture, before writing the result:
1) ask no more than 3 clarifying questions about the priority, constraints and who will read the result;
2) propose 3 approach options with "what we gain / what we lose";
3) wait for my choice and a short rationale;
4) only then prepare the final document or plan.
Skip this sequence for small, unambiguous tasks.
Resources
- Paper: Structured pre-generation elicitation versus single-shot prompting in AI-assisted enterprise decision-making: a randomised online experiment (working paper).
- Author: William Scott-Jackson, Oxford Centre for Impact Research (OCIR). The prototype scaffolding layer is called Cognistance.
- Key works cited in the paper: "cognitive surrender" (Shaw & Nave), "workslop" (Niederhoffer et al.), cognitive forcing functions (Buçinca et al.), metacognitive demands of generative AI (Tankelevitch et al.), the consultant experiment (Dell'Acqua et al.).