The Prompt Is Not the Hard Part

You have seen the posts. Probably several this week.

Use these 12 prompts to get McKinsey grade strategy for your business. This one prompt replaces a $200,000 consulting engagement. I asked ChatGPT to build my three year plan and the output was better than what our consultants delivered.

The comments are full of people asking for the prompt pack. There is always a way to buy it. A gated download in exchange for your email, a paid newsletter, a $49 bundle, a cohort course starting Monday. The prompts are the product now, and business strategy is where they sell best, because it is the one thing every business operator knows they are supposed to have.

Here is the part most responses to this get wrong. The prompts work.

Feed a good model a well constructed prompt with a bit of context about your business and you will get back a coherent situation assessment, a defensible set of strategic options, a reasonable evaluation of each, and a recommendation with an implementation outline. It will be structured properly. It will use the frameworks correctly. Sometimes it will even be more rigorous than what gets presented in executive conversations today.

I have spent twenty years delivering strategy engagements and I am not going to pretend the prompts don't work as intended. If your reaction to these prompt packs is that the output is shallow, go and try a few of them properly. The results are not shallow.

So let me say the uncomfortable part out loud, because it is the whole point of this piece.

Generating the strategy is never the hard part. It has never been the hard part. And now that it is basically free, we are all about to find out where the hard part truly hides.

What you actually got

Look carefully at what came out of that prompt. You got an answer.

An answer is a real thing and it has value. But strategy is not an answer, it is a system that has to survive contact with an organization for years. And there are three things a chat window structurally cannot give you, no matter how good the prompt is.

It does not know your company's core challenges

You pasted in some context. Revenue, headcount, a paragraph about your market, maybe last year's plan.

Now think about what you did not paste in. The initiative you tried in 2023 that failed, and the specific reason it failed, which everyone who was there remembers and nobody wrote down. The fact that your ops lead and your CFO have irreconcilable views on where margin actually comes from. The client concentration that makes one of those "growth options" much riskier than it reads on paper. The three business cases already approved that quietly assume conflicting value generation from the same headcount.

You cannot paste that in. Not because you did not think of it, but because your core organizational challenges cannot be distilled into a prompt.

That is not a complaint about convenience. When MIT's NANDA initiative looked at enterprise AI last year it found that about 95 percent of pilots delivered no measurable return, and the reason it landed on was not model quality. The report called it the learning gap. The tools do not retain feedback, do not adapt to context, and do not improve over time. Every prompt starts from zero, so the organization never gets any further ahead than the last person to write a good one.

The output looks like it is about your company. When it's actually about a generic company that resembles yours in the four dimensions you had time to describe. A human will always be needed to close that gap. Ideating the one initiative only your company could run is what separates your strategy from a generic one, and no prompt will get you there.

It does not include anyone else

The answer is in your chat window and not shared with the rest of your organization.

Strategy is a coordination problem before it is an analysis problem. Its entire job is to get a large number of people to make aligned decisions when you are not in the room. An answer only you can see does not do that, and it cannot be argued with, which matters more than it sounds. A strategy nobody challenged is not a strategy that survived scrutiny. It is one that never faced any.

Then the second person needs the same analysis. They open their own chat, write their own prompt, and get their own slightly different answer, because that is how these models work. Now you have two. Give it a quarter and you have nine, all reasonable, all subtly different, and the organization spends its meetings reconciling versions instead of deciding anything.

I wrote about this at more length, because it turns out to be the expensive part. The tool is governed now. The output is what fell into shadow.

It is confident, not calibrated

This is the one people find hardest to hear, because confidence reads exactly like competence.

When the model tells you an initiative is high risk, or that an option is likely to deliver, it is producing the words that best fit the pattern of the question. It is not drawing on whether its previous hundred recommendations actually worked out, because it has no idea. Nothing feeds back. There is no loop. There is no predictive model behind a generative chat.

Researchers at OpenAI have been blunt about where that comes from. Their argument is that training and evaluation reward guessing over acknowledging uncertainty, because a model that produces a confident answer scores better on the benchmarks than one that says it does not know. The confidence in the writing is not a signal about accuracy. It is a signal about what the scoring rewarded.

A calibrated prediction is a much different output. It comes from a model that has been trained on outcomes and tested against held out data, so that when it says seventy percent it is right about seventy percent of the time. You can put a calibrated number in front of a board and defend how it was derived. A confident sentence is not that, however well written it is.

Both sound the same in a document. Only one of them should influence a decision.

The bottleneck was never the deck

Here is the thing the prompt pack economy has got backwards.

Think about the strategies you have watched fail. Not the ones that were doomed from the start, the ones that were genuinely good and still went nowhere. Ask yourself honestly what went wrong.

It was almost never that the thinking was not smart enough.

Donald Sull and his colleagues went looking for the answer and reported in Harvard Business Review that two thirds to three quarters of large organizations struggle to execute. What they found breaking was not alignment up and down the chain, which is where most executives look first. It was coordination across units, and the ability to adapt once the plan met the year it was written for.

This has been my experience as well. The large three year strategy was developed over a week with every top executive in a conference room. Distilled into many documents, refined over a couple of months, then published to the rest of the organization. Division leaders like me were then assigned across towers to execute. But none of us were in the room when the thinking was baked, when the assumptions were challenged, and once we each defined our programs to execute, we had no way to align with the original thinking. We could only align with the distilled output documents.

None of that is a strategy generation problem. All of it is an execution and memory problem, and a better prompt makes none of it better.

What free strategy generation actually does is remove the last excuse. When the analysis was expensive, "we have not had the capacity to revisit the plan" was a real answer. It is not any more. The plan can be revisited on Tuesday for the price of a coffee. So the question moves, immediately and uncomfortably, to whether your organization can act on it.

For most companies that answer is no, and the reason is not intelligence. It is that there is no living system holding any of it.

What actually closes the gap

I am building Analyzt AI, so treat the rest of this as interested. But the gap above is the thing we built against, so it is worth being concrete about what a system does that a chat window cannot.

None of this replaces the person doing the thinking. The ingenuity stays where it always was. What changes is everything underneath it, because that is the part that has been failing.

It knows your organization, and it keeps knowing it. Your strategies, business cases, risks, initiatives and outcomes are structured objects that persist. When you run an analysis it is grounded in what is actually there. You are not describing your company from scratch every time. The context compounds instead of resetting.

It is shared by default. Run an assessment and it belongs to the organization, not to your chat history. The next person reads it, argues with it, and builds on it rather than buying their own slightly different version. Run a discovery chat session and share it across your organization directly in Analyzt AI.

The predictions are calibrated. Fourteen custom machine learning models score every strategy, outcome, initiative and business case. Predictive models, not generative ones. They are trained on outcomes and they can tell you how confident they are and why. All ML/AI outputs in Analyzt AI are wrapped with an explanation a human can understand, what the data you are seeing implies, why it's important, how to make decisions from it, and why you should question it. Which is never something you'll get out of a well written paragraph from chat.

It watches what happens next. Promised benefits get tracked against realized ones. The status pack rebuilds itself from live data every period. The plan does not decay quietly between board meetings, because keeping it current costs nothing.

Analyzt AI is methodology neutral. OKR, Balanced Scorecard, Hoshin, SWOT, Porter's Five Forces, Blue Ocean, PEST, McKinsey 7S. Your method is your method. We are not selling you one we picked or created. Which is more than can be said for the prompt pack, since what those actually encode is one firm's framework with the firm removed.

So go and use the prompts

I mean that. Run them. They are genuinely useful, they will sharpen your thinking, and anyone telling you the output is worthless is protecting something.

Just be clear about what you have when you finish. You have an answer, produced by something that does not know your company, cannot show anyone else, and has no idea whether it has ever been right before.

The hard part starts after that, and it always did. We just used to be able to hide it behind how long the analysis took.

---

If any of this sounds like your organization, the beta is open and free. And if you run a consulting firm working through what this means for how you deliver, I would like to compare notes.