Show, Don't Tell: Examples Beat Adjectives
A working note on why describing the output you want works worse than showing it.
When students want a particular kind of output, the reflex is to describe it: pile on adjectives until the description feels precise. But a stack of adjectives is the vaguest instruction you can give. The more reliable move is to stop describing and start showing.
The common move
Every adjective feels like it sharpens the target. It feels like the more you specify, the closer the model gets to the thing in your head. The question worth asking: what is the model actually doing with those words?
What you think it does, and what it actually does
- Transmits a precise target. You picture a specific voice and assume the adjectives carry it across.
- Narrows with each word. More adjectives feel like tighter aim.
- Specifies the result. You believe you've told the model what to produce.
- Averages every meaning. The model blends every "witty" and "professional" it has seen, landing on the blandest center of all of them.
- Compounds the vagueness. Each added adjective is one more thing to average, not one more constraint.
- Gestures, doesn't specify. You've described a direction, not shown a destination.
Adjectives are abstractions, and the model resolves an abstraction by averaging. That's why "witty and engaging" so often comes back as generic: the average of everything is nothing in particular.
The more honest move
Show the model two or three examples of the output you want. It will infer the pattern far more precisely than any adjective could describe it.
This technique has a name: few-shot prompting, giving the model a few examples (shots) of the target before asking for the real thing. It works because a demonstration carries information a description can't. "Witty" is an abstraction the model has to guess at; a sentence you actually find witty is a concrete pattern it can match.
Try this
Next time you're stacking adjectives, stop and ask: can I show an example instead?
If you have one, paste it and ask the model to match it. Two or three examples beat any list of descriptors.
If you don't have one, that's worth noticing too. It may mean you don't yet know the target concretely enough to recognize it, which no prompt can fix for you.
The principle underneath
The model can't read the picture in your head; it can only pattern-match what you put in front of it. Show it the target instead of describing it.