Why writing prompts is quietly hurting your results
Last month we checked our internal collection of prompt examples and found instructions going past two thousand words. We thought more details meant results but our tests showed the opposite effect.
When directions get long models start missing important formatting rules hidden in the middle sections. Our output became more consistent only after we made those big sections shorter and used bullet points instead.
We learned to see prompts as code functions not as assignments. Keep the focus small explain input values clearly. Take out all the friendly conversation words.
A short instruction set makes the model use its knowledge instead of getting lost in different style rules. Fewer words often mean thinking.
Do you like system prompts or small separate instructions, for hard tasks?
Replies
Small separate calls, every time, for anything that runs in production. In my song app the lyrics, the title, the cover brief and the rescue of a request the music model refused are each their own call with their own short prompt. The one that names the song and briefs the cover artist is about ten lines. The other half of it is a response schema: the format rules live in the schema, not in the prose, so they cannot get lost in the middle because they are not in the middle at all. Prose for judgment, schema for shape. And a separate call is something you can test and swap on its own, a two thousand word prompt is one big function with no tests.