Most "best ChatGPT prompts" lists are 200 templates you will never use. This is a shorter list of prompts that hold up, plus the pattern underneath them so you can write your own.
The pattern
Almost every good prompt has four parts:
- The material — paste it, don't describe it
- The audience — who reads the output
- The shape — length, format, structure
- The bans — what you don't want
Miss any one and you get the model's default, which is longer, more structured and more generic than you wanted.
What does not help: elaborate role-play ("you are a world-class expert with 20 years' experience"), politeness, threats, or offering it money. These circulate widely and do close to nothing.
Writing and editing
Tighten a draft
Cut this by a third without losing any of the specifics. Keep the opening line. No closing summary. (paste)
Match your voice
Here are three things I've written. Match this voice — notice the sentence length and how little jargon there is. Now write [X]. (paste 3 samples)
Examples work; adjectives like "professional but friendly" do not.
Change the audience
Rewrite this for someone who runs a small business and isn't technical. Same facts, no jargon, shorter sentences. (paste)
Kill the AI tells
Rewrite without: tricolons, em-dash asides, "in today's", "it's not just X it's Y", and any closing paragraph that restates the opening.
Thinking and deciding
Steelman the other side
I'm leaning towards A over B for [situation]. Make the strongest possible case for B, then tell me which of my reasons for A is weakest.
Find the hole
What's the weakest part of this argument? Name the specific sentence and explain why. Don't be polite about it. (paste)
Pre-mortem
It's six months from now and [plan] has failed. Write the three most likely explanations, most likely first.
Surface the assumption
What is this plan assuming that might not be true? (paste)
Reading and extracting
Action items from a transcript
Pull out every action item, who owns it, and any deadline. Table format. If an owner isn't stated, write "unassigned" — don't guess. (paste)
The "don't guess" clause matters. Without it, models fill gaps.
Compress a long document
Summarise in under 200 words for someone who has to make a decision about it. Lead with what they need to decide, not with background. (paste)
Answer from the document only
Answer only from the text below. If the answer isn't in it, say "not covered". (paste)
This is the single most useful anti-hallucination instruction there is.
Learning
Explain at the right level
Explain [topic] to someone who knows the general area but not this specific thing. Skip the basics. One concrete example.
Find what you're missing
I understand [X] as [your explanation]. What am I getting wrong or leaving out?
Test yourself
Ask me five questions about [topic], one at a time, that would reveal whether I actually understand it. Wait for my answer before the next one.
Code
Reviewable diff
Show me only the lines that change, not the whole file. (paste code and the problem)
Debug with context
Here's the error in full and the function it's thrown from. I've already checked [X] and [Y]. (paste)
Three habits worth more than any prompt
Follow up instead of restarting. "Shorter." "Too formal." "Keep paragraph two, redo the rest." Starting over throws away everything it learned about your task.
Paste, don't describe. The most common reason for a generic answer is that the model was given a generic question.
Ban things explicitly. Negative instructions work well and are badly underused.
Common questions
What makes a good ChatGPT prompt? Your actual material, a named audience, a specified length and format, and an explicit list of what to avoid.
Do role-play prompts work? Barely. "You are an expert X" is far less effective than concrete instructions about audience, length and constraints.
Should I be polite to it? It makes no meaningful difference to output quality. Do whatever you prefer.
How do I stop it making things up? "Answer only from the text below; if it isn't there, say so." Grounding it in material you provide is the only reliable method.
Why do I get different answers to the same prompt? These models sample rather than always picking the most likely next word. Same prompt, different answer, by design.
Do prompts transfer between models? Mostly. The four-part pattern works everywhere; specific phrasings tuned to one model usually do not matter as much as people think.