Prompts That Hold Up
A prompt is a work order, not a wish. The four parts that make the difference – and why one example beats three adjectives.
Boaz Lichtenstein
Most bad answers are good answers to a different question. Type “write me a professional text about our new product” and you have left three decisions open: for whom, in what form, based on what. The model then makes them for you – and statistically gets them wrong.
A prompt that holds up is therefore not a wish but a work order. It has four parts.
1. The task, in one sentence
One task per prompt. “Summarise and translate and shorten” is three tasks, and a model does them with differing quality. Work sequentially instead – summarise, check the result, then translate – and you get a chance to intervene at every step.
2. The material
The biggest lever of all: give the model the thing it should work from. Your own text, the numbers, the customer’s email. Without material a model writes the most probable thing – which is the most generic thing. With material it writes about your case. Why that reaches deeper than any art of phrasing is the subject of the next station on this path.
3. The form
Not “short”, but “five bullet points, 15 words each at most”. Not “a table”, but the columns it should contain. Numbers can be checked, adjectives cannot – and what can be checked is what you get.
The strongest input here is an example: one paragraph showing what the result should look like. An example beats three adjectives because it needs no interpretation.
4. The boundary
Say what must not happen. “Invent no numbers; if a figure is missing, write MISSING.” That single line turns a hallucination into a gap you can see. It is the cheapest quality gain available when working with language models.
When the answer does not fit
The usual reflex is to rephrase the prompt. Usually, though, the problem is not the phrasing but a missing input. The more useful question is: what would a new colleague have needed to get this right? That is exactly what the model is missing too – context, an example, a boundary.
Two habits that pay off: collect the prompts that work somewhere, instead of reinventing them each time. And split long briefs into two steps, because then you can correct in the middle instead of discarding everything at the end.