A prompt with no examples is a guess about what the model will do

A founder at a legal-tech startup showed me a prompt he was frustrated with. It was long. It was careful. It described exactly what he wanted the model to produce, in detailed prose, and the model kept producing something adjacent to it but not quite it. The structure was almost right. The tone drifted. The output surprised him more often than it should have.

He kept adding paragraphs. The surprises kept coming.

Prose describes, examples demonstrate

His prompt was all instruction and no example. It told the model what the output should be, in sentences. “Produce a summary with the key points, in a neutral tone, organized by topic, with a short conclusion.” Every word of that was reasonable, and none of it pinned down the actual shape. What does “organized by topic” look like exactly? How long is “short”? What does his idea of neutral sound like? The prose gestured at an answer and left the model to fill in the specifics, which it did, differently each time.

We added three examples. Real input, and the exact output he wanted next to it. Not a longer description of the output, the output itself. The next run came back in the shape he had been trying to describe for a week. The examples did in three samples what another paragraph of instruction had failed to do, because they showed the model the target instead of talking around it.

He had been trying to specify a shape in words. The model needed to see the shape.

Examples pin down what prose leaves open

Instructions are good at intent and bad at precision. They tell the model what you care about. They struggle to convey the exact format, the length, the register, the thousand small choices that make an output feel right. A model can read “be concise” a hundred ways. It reads one concrete example of concise and knows what you mean.

This is why an instruction-only prompt so often surprises you. You have described a region of possible outputs and left the model to pick a point in it. Every run picks a slightly different point. Examples collapse that region. They show the model where in the space you actually want to land, and the output stops wandering.

Getting the benefit is a matter of choosing examples with care.

Show, do not only tell. For any output shape that matters, include an example of it rather than another sentence describing it.
Pick examples that carry the format. The example is where length, structure, and tone get fixed, so choose ones that demonstrate exactly those.
Include a hard case. An example of a tricky input handled well teaches more than a paragraph warning about it.
Prefer a few good examples over more prose. When output drifts, reach for an example before another instruction.
Keep examples current. They set the behavior, so a stale example teaches a stale format.

The move is simple. When words are not landing the behavior, stop writing more words and show the model what you want.

How we approach it at Density Labs

In the AI Opportunity Assessment, our fixed two-week engagement priced at $2,500, an instruction-only prompt that keeps surprising the team is a familiar pattern. The fix is usually not more instruction. It is a small set of well-chosen examples that show the model the exact output, including the format and tone the prose could not quite capture. We pick those examples against a real sample so they demonstrate the cases that actually occur.

A prompt with no examples leaves the output to chance and then acts surprised by the result. Show the model what you want, and it will stop guessing.

Words tell the model your intent. Examples tell it your answer.