"Just ask it differently."
You've heard this. Maybe you've said it. It's the standard advice when an LLM gives you a bad answer. And it's correct. But nobody explains what "differently" actually means.
Here's what changes when you improve a prompt.
Vague prompt vs. precise prompt
Most people write prompts like this: "Summarise this article."
And they get a generic summary that could apply to anything. Then they try again. And again. And wonder why the output is always mediocre.
The issue isn't the model. It's the context. You gave it almost nothing to work with.
Here's the same request, improved: "Summarise this article in three bullet points. The audience is a non-technical marketing manager. Focus on the business impact, not the technical details. Keep each bullet under 20 words."
Same model. Same article. Completely different output.
Four things that always improve a prompt
- Give it a role. "You are a senior copywriter with 10 years of B2B experience" gets very different output than no role at all. You're telling it which patterns to predict from.
- Give it the format. "Three bullet points", "one paragraph", "a table with two columns". Don't make it guess what you want.
- Give it constraints. "No jargon", "under 100 words", "don't mention pricing". Constraints narrow down the output space dramatically.
- Give it an example. Show it what a good answer looks like. It will pattern-match against that. This is the most underused technique and often the most powerful one.
The thing most people skip: telling it what to avoid
A prompt that says "write a LinkedIn post about our product launch" will give you something generic. A prompt that says "write a LinkedIn post about our product launch, no corporate language, no bullet points, don't start with a question" cuts out half the bad outputs before they happen.
Negative instructions are just as important as positive ones.
Prompting is iteration, not magic
The best prompts aren't written in one go (I know you probably hate to hear this). You start with something reasonable, look at the output, figure out what's off, and add more precision. Every iteration narrows the pattern.
Think of it less like giving an order and more like briefing a colleague who just joined the team. The more context you give upfront, the less back-and-forth you need.
Next week: what is RAG, and why does it matter when you want an LLM to actually know things?