How to Write Good Prompts: My Notes After 6 Months of Use
After using LLMs dozens of times a day at work, I noticed patterns emerging before I knew it. Nothing grandiose—just following these alone makes a big difference in output quality.
- Put the role and goal in the first paragraph. Start with "You are an expert in X, and I want Y."
- Show the desired output format with an example. One example is far stronger than explaining it in words.
- Explicitly state what not to do. A single line like "Don't guess—if you don't know, say you don't know" greatly reduces hallucinations.
- Don't ask for everything at once; break it into steps. The three stages—draft → critique → revise—worked best for me.
The last three-stage process especially was a game changer. At first I just said "write it well," but now it's become a habit to have the model critique its own answer.
I'm sure it varies by person, but if you have other tips, please share.
8 answers
Agreed, especially with step 3. Once you get a draft and then tell it to "point out just three weaknesses in this answer," it comes out like someone else wrote it.
In short: just specifying the role and format alone gets you halfway there. The rest is a matter of preference.
Oh, I didn’t know this. I hadn’t thought about the critique step haha I’ll try it right away tomorrow.
For me, running "Rewrite this answer as if you're explaining it to an elementary school student" one more time at the end cleared out all the fluff, which I really liked. It feels like adding one extra step to the three-step process.
Got a source? Have you ever confirmed that "hallucinations drop significantly" with numbers rather than just a gut feeling? I'm curious about the paper or benchmark evidence.
Splitting it into steps is a good idea, but doesn’t running 3 turns every time waste tokens and time? It feels like overkill for simple tasks. That seems like something that should be decided based on task difficulty.
One practical tip to add: saving your prompts as files and version-controlling them makes a huge difference. If you keep writing new ones from scratch every time, you forget what actually worked. I collect them by case in Notion, and that alone cut my rework time by more than half.
Well, isn't the idea that "telling it not to guess reduces hallucinations" a bit outdated? These days, models often get things wrong in similar ways whether that prompt is there or not. If anything, I feel like including the reference documents makes a much bigger difference in practice.