Want better outputs from LLMs? Forget about data and prompts.

A brilliant prompt can be in the wrong place

You can spend an extraordinary amount of time writing the perfect prompt for brand positioning. Define the role, specify the framework, add examples, explain what good looks like and list the clichés to avoid. The output will probably improve.

There is a more basic question: should you be developing the positioning yet? Perhaps the audience is not understood, the competitive analysis is superficial or nobody has agreed which problem the brand needs to solve. A better positioning prompt helps you answer the wrong question more efficiently.

Sequence matters because strategy is a collection of tasks performed in an order where one result changes what should happen next. That is difficult to solve with a clever prompt.

More context has the same problem

The other instinct is to give the model everything: research, previous presentations, workshop outputs, brand guidelines, campaign results, audience studies and meeting notes.

This may mean giving an LLM a beautifully organised pile containing things that are true, things that used to be true, things somebody suggested once, things the client likes, things everybody knows are nonsense and a 96-page presentation with the important bit on slide three.

Humans do not treat this information equally. Someone who has worked on the client knows which positioning was superseded and which workshop ideas went nowhere. That hierarchy rarely exists in the documents. To the model, context without hierarchy can become a large collection of things that look equally worthy of attention.

We have seen this movie before

The enthusiasm for feeding more information into AI reminds me of the data-lake enthusiasm of the 2010s. Companies collected enormous amounts of data because data was valuable. Then many discovered that owning a lot of data and knowing what to do with it were different capabilities.

We risk repeating the mistake with context. Give the model everything. Make the context window bigger. Connect another database. Upload another document. None of these actions answers the important question: which information matters for the decision we are making now?

That is a process question.

Good strategic work has dependencies

Consider one sequence: brief, audience, positioning, campaign strategy. The brief diagnoses a consideration problem. The audience work then shows that people know the brand and distrust its claim. That finding should change the positioning task. If it does not, the team may polish the third prompt while leaving the dependency between steps two and three broken.

If every step begins as a fresh conversation, the user has to copy the audience profile, paste the positioning, explain the client, remember which conclusions mattered and correct old assumptions. The human is acting as middleware.

Improving that dependency eventually matters more than squeezing a little more performance from prompt three. The audience conclusion should enter positioning with its status and evidence intact. If the positioning changes, campaign strategy should know that the previous route is obsolete.

Do not jump ahead

LLMs are extremely accommodating. Ask for a campaign strategy and you will probably get one. The model rarely puts its feet on the table and says your brief is contradictory, your audience definition is useless and you have not decided what problem you are solving.

Experienced strategists are often more annoying. They ask the uncomfortable question first, notice that three objectives cannot all be priorities and sometimes refuse to solve the requested thing until the problem underneath it makes sense.

That behaviour can be encoded. A process can require diagnosis before solution, surface ambiguity and carry the result into the next stage. This requires deciding how good work should happen.

Process includes evaluation loops

Sequence alone is not enough. Useful workflows include a loop: generate, evaluate, challenge, revise and commit. A first positioning route may satisfy the brief and fail the competitor test. A campaign strategy may be coherent and produce lifeless creative work. The process needs somewhere for that evidence to travel backward.

Commitment matters too. After challenge and revision, somebody has to choose a route and make its status clear. Otherwise the system creates a well-organised collection of options while the actual strategic decision remains unresolved.

The prompt still matters

A terrible prompt inside a brilliant process remains a terrible prompt. Models need clear instructions, relevant context and criteria for a useful result. I am arguing about where the next unit of effort goes.

Once the prompt is competent and the information is good enough, polishing the instruction for the seventeenth time produces diminishing returns. Larger gains come from asking whether the model is doing the right task at the right moment, with the right information, informed by the right things that happened before.

When people become very good at a job, much of their process disappears from view. Years of experience, discarded directions and absorbed research sit behind the moment when something clicks. The spark and the machinery behind it are both real. If you want consistently better output, study how the work gets done when a very good human does it, then encode that.