Prompting Is Specification
When the output is wrong, was the model wrong or was the question?
The idea
A model can only work from what the prompt supplies, so vague requests are answered by assumption.
In the real world
Asking for a summary without saying for whom produces one aimed at nobody.
Going deeper
A model fills every gap you leave. Ask it to make something better and it must silently choose a criterion — shorter, warmer, more formal, more rigorous — and the disappointment when it chooses differently from your unstated intent is not really a model failure.
The fields worth stating explicitly are the audience, the format, the length, and what to leave out. The last is the most neglected and often the most useful, because a model has no way to know which of the many true things it could say are irrelevant here. Treating the prompt as a brief for a competent stranger gets you most of the way.
Where it stops applying
Over-specifying can be worse than under-specifying for open-ended or creative work, where the value is in options you had not considered. Constrain tightly when you know what you want, loosely when you are exploring.
Why it matters
It moves the problem from the model being wrong to the request being under-specified.
Try this today
Rewrite one prompt to name the audience, the format, and what to leave out.
Test yourself
Someone asks a model to make this better and dislikes the result. What information did the model have to invent?
Show the answer
Better according to whom, and by which measure: shorter, more formal, more persuasive, more accurate. With no criterion supplied it selects a plausible one, and the disappointment is that it chose differently from the unstated intent.
Learn this in the feed Answering from memory, then again days later, is what makes it stick.