Using Prompts Like This Makes Results Completely Different (In Practice)
I use models almost every day at work to draft planning documents or summarize reports. At first I threw in things like “write this well,” but I kept getting the same bland results, so after a few months of use I’ve kept only the methods that actually work.
- Lock in the role and audience first
If you include something like “You are a B2B SaaS product planner. The reader is a decision-maker,” the tone gets fixed right away. Expressions like “like an expert” had almost no effect.
- Show the desired output format with an example
Attaching one short sample is far more accurate than explaining it in words. If you want a table, it’s better to include even an empty table.
- Don’t ask for everything at once
“Research, organize, summarize, and polish” always makes the middle steps sloppy. Splitting it into three stages—research → structure → polish sentences—makes a different result.
- Specify prohibited conditions
Just adding these three lines—“Do not speculate, do not use unsupported numbers, and if you don’t know, say you don’t know”—greatly reduces plausible-sounding lies. This had the biggest noticeable impact.
- State length as numbers
“Short” is almost ignored. If you give numbers like “three paragraphs, each no more than three sentences,” it follows well.
Of course, things change slightly when the model changes. Still, I’ve been using these five for almost a year, and they’ve never failed me. I’d also love to hear other people’s tips.
8 answers
I really relate to the prohibitions... Just adding the one line “If you don’t know, say you don’t know” makes the output feel much more trustworthy.
Well, role assignment varies a lot depending on the model/version, so I’m not sure it makes that much of a difference. For me, just including examples had the biggest effect.
Got a source?
I’ve also been splitting it into three steps for a while now, and the middle definitely feels less awkward. But the key is to paste the previous step’s output back into the prompt when moving to the next step. Otherwise, the model forgets what it did earlier and ends up just as bland.
Agreed, but isn't including table examples a waste of tokens?