Vague in, vague out

Models can't read your mind. When you leave a choice unspecified — length, audience, tone, format, what to avoid — the model fills the gap with the most generic option, because generic is the safest bet across everything it has seen. "Write about dogs" could mean a poem, a vet-bill explainer, or a 2,000-word history, so you get bland, middle-of-the-road text. The fix isn't a better adjective; it's removing the choices you don't want it to make.

Constraints are a feature, not a cage

It feels backwards, but adding constraints usually makes answers better, not narrower. Each constraint you supply is one less thing the model has to guess. Compare:

• Vague: "Write about adopting dogs."
• Specific: "Write a 150-word intro about adopting senior dogs, aimed at first-time owners, warm but not sappy, ending with a call to visit a local shelter."

The specific version fixes the length, subject, angle, audience, tone, and ending. There's almost nothing left to get wrong.

The dials worth setting

When a task matters, decide these on purpose instead of leaving them to chance:

• Audience — "explain to a beginner" vs. "for an expert" changes everything.
• Length — "in one sentence," "under 100 words," "a thorough guide."
• Tone — formal, playful, blunt, encouraging.
• Format — prose, bullets, a table, JSON, a step-by-step.
• Boundaries — "no jargon," "no emojis," "don't invent statistics."

Show an example when words fail

Some things are easier to show than describe. If you want a very particular style, paste one example of it. A single good example often replaces a paragraph of fussy instructions — the model simply matches the pattern it sees.

Specificity is the cheapest upgrade in prompting. You're not limiting the model; you're handing it the answer key to what you consider good.

Try this: Write the vaguest possible version of a real request ("help me with my resume"). Then rewrite it with five constraints — role, audience, length, tone, and what to avoid — and run both. The gap between the two answers is the whole skill of prompting in one experiment.