Summary
Before asking whether a model is smart enough, ask whether you've set it up to perform. Promoting the model means framing the request with enough context, defining what quality looks like, clarifying the role it should take, and giving it structure to reason within. This piece covers why under-specifying the task under-promotes the model, why prompt optimisation is leverage rather than fussiness, and how the meta-prompt, asking the model to critique your prompt first, opens up a different tier of performance. When the output is weak, the problem is usually the prompt.
We talk constantly about how to use language models: getting better answers, building better workflows, automating the tedious parts. A more basic question tends to slip past: are we promoting the model correctly?
Not in the marketing sense, but in a more practical one. Are we setting the model up to perform at its best? Are we asking in the right way, with the right amount of guidance? Are we framing the interaction so it knows what success looks like?
Before asking whether the model is smart enough, it's worth asking whether the prompt is any good.
Models Don't Autocomplete Your Thoughts
Most people prompt as though it were a more capable Google search: type a phrase, get an answer. Modern LLMs behave far more like a collaborator responding to the structure, tone and clarity of what you give them.
When your prompt is vague, the model isn't failing. You're under-specifying the task. And under-specifying the task under-promotes the model, because how well it performs depends heavily on which capabilities you actually activate.
What Promoting the Model Means
In practice it means five things:
- Frame the request with enough context for the model to reason properly
- Say what quality looks like: style, structure, tone, constraints
- Clarify the role it should take, whether strategist, tutor, analyst or editor
- Give examples when the task is unusual or the format matters
- Define the objective, not only the task
Skip these and you invite generic output. Do them well and the model behaves like a domain expert who has been properly briefed.
Prompt Optimisation Is Leverage
People skip this because it feels tedious. But prompting well isn't about adding complexity. It's about adding intent.
Unoptimised: "Explain reinforcement learning."
Promoted: "Explain reinforcement learning to a reader with basic Python experience but no ML background. Use one concrete analogy and include a simple code example. Keep the tone conversational."
Same model, two completely different outputs. The second version puts the model into a specific teaching role and sets the constraints that make clarity possible.
This isn't gaming the model. It's removing friction.
Your Prompt Is Part of the Output
It's tempting to treat the prompt as pure input. In practice it shapes the output directly. A well-built prompt communicates what context matters, what success looks like, what to avoid, what format serves the reader, and what the model should assume.
The shift is from telling the model what to do, to setting it up to succeed.
The Meta-Prompt
One of the most underrated moves is asking the model to evaluate your prompt before answering it.
"Before answering, analyse whether my prompt is well-defined, and suggest improvements if needed."
That creates a feedback loop and gives the model room to sharpen your thinking before it produces anything. You're no longer just promoting the model. You're improving your own ability to work with it, which is where the durable leverage sits.
The Short Version
Language models are responsive rather than psychic. A thin prompt gets you a thin answer; a well-structured one gets you something you can actually use.
So before criticising an output, ask the prior question: did I frame the task in a way that let the model show what it can do?
If not, the problem isn't the model. It's the prompt.