Name the Problem Before You Build the Answer
AI keeps handing back generic strategy for one reason: the goal it was given was generic. The work of defining the problem before using AI is the work that decides everything downstream. A model is shaped by how the problem is posed, not by how hard it is, so a loosely framed question returns the average of everything the model has read. The fix is not a longer, cleverer prompt. It is naming the uncomfortable truth about what you are actually solving, then turning that into a sharp goal. The moment you do, the machine stops being a vending machine and starts being a thinking partner. This is the Mirror Principle: if the output is generic, the reasoning was generic. Goal definition with AI, not prompt mechanics, is the lever. Name the problem, and the answer changes shape. The hard part was never the answer. It was the question.
Generic strategy back means a generic goal in
A senior leader opens a chat, describes a situation in a few competent sentences, and asks the model for a strategy. What comes back is fluent, well-structured and forgettable: the kind of advice that would fit any company in the sector, on any day, facing any version of the problem. The leader concludes the tool is shallow and moves on.
The tool is not the issue. The question was. There is now hard evidence for what most users feel: a large language model is shaped, in its strengths and its failures, by how the problem is posed to it, far more than by how difficult the problem is. Researchers at Princeton mapped this directly and named it the embers of autoregression: the same task framed two ways produces sharply different output quality, because the model is reaching for what is probable given the framing, not what is correct given the world (McCoy et al., PNAS, 2024). Pose the problem in the average way and you get the average answer, articulately.
We have a name for the diagnosis: the Mirror Principle. If the output is generic, the reasoning behind your question was generic. The machine is a mirror with a vocabulary. It returns the shape of what you put in, and a vague goal has a vague shape. So the first move when AI disappoints is never to scold the model or to reach for a cleverer prompt. It is to look at the question and ask whether it actually names what you are solving for.
The uncomfortable truth you have not named
Most strategy questions arrive pre-laundered. By the time a leader types them, the room has already agreed on the comfortable version of the problem, and the question carries that agreement in its phrasing. "How do we grow the category" sounds like a goal. It is a euphemism. Underneath it sits a truth nobody has said out loud: the category is not growing, the product is not the one buyers reach for first, and the honest question is why anyone would choose it at all.
Naming that truth is uncomfortable, which is exactly why it gets skipped, and skipping it is what produces the generic answer. The research on hard problems is unambiguous here: jumping straight to solving caps the quality of every solution that follows, and the move that pays most is to frame, and then reframe, the problem before touching the answer (Binder and Watkins, Harvard Business Review, 2024). The reframe is not a softer version of the work. It is the work.
This is what the Goal line of the 4-Lines is for. Of the four (Persona, Goal, the Flip, Sequence), the Goal is the one most leaders rush, and it is the one that decides whether the output is worth reading. A precise goal states the real constraint, the real stake, and what would have to be true for the assumed answer to be wrong. It is the difference between asking a model to write a strategy and asking it to solve the thing you have been avoiding.
The wrong problem
"How do we grow the category and win share back?"
Competes on the metric the leader already owns. Invites a generic plan to do more of the same, harder.
The right problem
"For whom are we the obvious fit the leader can never be?"
Shifts the basis of comparison from the metric we lose on to the fit we win on. Invites a defensible, specific answer.
Longer prompts are the reflex; a clearer goal is the lever
When the first answer disappoints, the common reflex is to reach for syntax. People add roles, constraints, formatting rules, few-shot examples and step-by-step instructions, building ever more elaborate prompts around a goal that is still vague. The prompt grows; the answer stays generic. The effort went to the wrong place.
The model was never waiting for cleverer phrasing. It was waiting for a clearer problem. Because output is shaped by how the problem is posed rather than by how hard it is (McCoy et al., 2024), a precise goal in plain language will out-perform a baroque prompt built around a loose one, reliably. This is why the 4-Lines spend their effort on the problem and not on the mechanics. Persona sets who is reasoning, Goal names what is actually being solved, the Flip makes the model interrogate before it answers, and Sequence orders the work. None of them is a prompt trick. Together they are a way of thinking about the question.
The longer bar is where most people spend. The shorter, red bar is where the answer actually changes.
None of this is an argument against precise instructions. It is an argument about order. Sharpen the goal first. Then, and only then, is it worth shaping how the model should work the problem. Reverse that order and you are decorating a vague question, which is the most expensive way to stay generic. For the deeper treatment of why this matters, the difference between the Havruta Methodology and prompt engineering sets it out in full.
Winning as the challenger by changing the question
Picture an invented case, deliberately generic. A marketing and strategy team is rebuilding the narrative for a product that is not the category leader. The instinct, and the question they bring to the model, is "how do we close the gap with the leader and win share". The model, doing exactly what it was asked, returns a competent plan to compete on the leader's terms: match the features, sharpen the pricing, out-spend where possible. Every line of it is generic, because the question accepted the leader's measure as the measure that matters.
The reframe is small and total. The challenger does not win on the dimension the leader owns; it wins on a dimension the leader cannot occupy without abandoning what makes it the leader. So the right problem is not "how do we beat them on their metric" but "what are we the obvious fit for, that they structurally cannot be". Asked the second question, grounded in the team's own customer reality, the model stops averaging and starts discriminating: it surfaces the segment that the leader serves badly, the use case the leader's scale makes awkward, the promise the leader would contradict itself to make.
That is the move shown in the diagram above. The wrong problem competes on the metric you lose on. The right problem shifts the basis of comparison to the fit you win on, and the output that follows is arguable, specific and defensible, not a forced fit and not a generic plan. Nothing about the model changed between the two answers. The question did. The reframe, again, is where the gain lives (Binder and Watkins, 2024), and the challenger's whole strategy turns on naming the right one.
The Flip, turned inward
The Flip is the move that defines the methodology: instead of instructing the model to answer, you instruct it to question you first, to interrogate the request before it serves a response. There is now direct evidence that this is not a stylistic preference but a mechanism that works. When researchers trained a model to model future conversation turns and ask a clarifying question before answering, it beat one-shot answering on the resulting decisions (Zhang et al., ICLR, 2025). A question asked before the answer is worth more than a better answer.
The discipline this essay argues for is that same Flip, turned inward. Before the model ever sees your goal, run the interrogation on yourself: what am I actually solving for, what is the uncomfortable truth under the comfortable phrasing, and is the question I am about to type the right question or the convenient one. Do that well and the model has little left to clarify, because the goal arrived already sharp. The leaders who get the most from AI are not the ones with the longest prompts. They are the ones who have already done the Flip on themselves before they open the chat.
This is the line between the two ways of working. Ask a vague question and you get a vending machine: a request in, a generic answer out. Name the real problem and you get a thinking partner that helps you reach a better question than the one you walked in with. The same model sits behind both. The difference is whether you named the problem first.
The machine is a mirror with a vocabulary. Name the wrong problem precisely and it will solve the wrong problem beautifully.
The discipline has a name
Reaching for a cleverer prompt is the default, and it is a trap that keeps the answer generic. Naming the real problem before you build the answer is a discipline, and it is learnable. We call it the Havruta Methodology, after the oldest form of rigorous study we have: two minds, one question, an argument that ends in a sharper version of both sides. Applied to AI, it puts the effort where it pays, on the problem rather than the syntax, and it turns the model from something that dispenses answers into something that helps you earn them.
So before your next strategy session with a model, do not start by writing the prompt. Start by naming, in one honest sentence, the thing you are actually trying to solve, including the part the room has been avoiding. Then ask. The quality of what comes back will tell you whether the problem was named. This habit is the heart of the Cognitive Pillar, and it is where every leader we work with starts.
If this resonates, two companion pieces go deeper: on how a leader should use AI in executive decision-making, and on the daily reasoning practice itself in the Havruta programme.
Frequently asked questions
Why does AI keep giving me generic strategy?
Because the goal you gave it was generic. A model is shaped by how the problem is posed, not by how hard the problem is (McCoy et al., 2024), so a loosely framed question returns the average of everything the model has read. This is the Mirror Principle: if the output is generic, the reasoning behind the question was generic. The fix is not a longer prompt. It is naming the real problem precisely, then asking the model to reason from that.
How do I define the problem before using AI?
Name the uncomfortable truth about what you are actually solving for, not the comfortable version everyone has agreed on. State the real constraint, the real stake, and what would have to be true for your assumed answer to be wrong. That sharp goal is the Goal line of the 4-Lines. Research on hard problems is consistent: jumping straight to solutions caps their quality, and the gain sits in framing and reframing the problem first (Binder and Watkins, 2024).
Is a longer, cleverer prompt the way to get better answers?
No. Reaching for longer, cleverer prompts is the common reflex and the wrong lever. The model is not waiting for more elaborate syntax; it is waiting for a clearer problem. The 4-Lines spend their effort on the problem (Persona, Goal, the Flip, Sequence), not on prompt mechanics. A precise goal in plain language will out-perform a baroque prompt built around a vague one, every time.
What is the Flip turned inward?
The Flip is the move of instructing AI to question you before it answers. Turned inward, you apply that same interrogation to your own goal before the model ever sees it: what am I really solving, and is the question I am about to ask the right one? Teaching a model to ask a clarifying question before answering measurably beats one-shot answering (Zhang et al., 2025); doing it to yourself first removes the need for the model to.
How do you reframe a problem to compete as a challenger?
Shift the basis of comparison. A challenger that is not the category leader loses on the metric the leader owns, so the work is to name the dimension on which it wins and make that the question. The wrong problem is "how do we beat the leader on their measure"; the right problem is "what do we fit that they cannot". Reframing the problem, not working harder on the original framing, is where the gain lives (Binder and Watkins, 2024).
What is the difference between a vending machine and a thinking partner?
A vending machine takes a request and dispenses a generic answer; a thinking partner interrogates the request first and helps you reach a better one. The difference is not the model, it is whether you named the real problem. Vending Machine vs Thinking Partner is the central diagnostic of the Havruta Methodology, and the lever that moves you from one to the other is goal definition, not prompt length.
References
- McCoy, R. T., Yao, S., Friedman, D., Hardy, M. D., Griffiths, T. L. "Embers of Autoregression Show How Large Language Models Are Shaped by the Problem They Are Trained to Solve." PNAS, 2024.
- Binder, J., & Watkins, M. D. "To Solve a Tough Problem, Reframe It." Harvard Business Review, 2024.
- Zhang, M. J. Q., Knox, W. B., Choi, E. "Modeling Future Conversation Turns to Teach LLMs to Ask Clarifying Questions." ICLR, 2025.