Every “500 best ChatGPT prompts” list has the same problem: nobody uses 500 prompts. After two years of using these tools for real work, I use maybe a dozen patterns. The rest were noise.
What follows is that dozen, with the reasoning attached. The reasoning matters more than the wording — once you see why a pattern works, you can build your own.
Why most prompts fail
A bad prompt is usually bad in one of three ways. It asks for something the model cannot verify, it gives no constraints, or it asks for too many things at once.
“Write me a marketing plan” fails all three. The model does not know your budget, your market or your product, so it invents a generic one. That is not the model being stupid. That is you asking a question with no answer.
1. Give it a role with a reason
“Act as a lawyer” is weak. “You are reviewing this contract for a freelancer who cannot afford a dispute — flag anything that could cost them money” is strong.
The role is not magic. What helps is the stake: it tells the model which details matter.
2. Show one example instead of describing ten rules
Describing a format takes a paragraph and still gets misread. Pasting one example of the output you want takes two lines and works.
This is the highest-leverage technique on the list and the one most people skip.
3. Ask for the reasoning before the answer
Adding Think through this step by step before answering genuinely changes output quality on anything involving logic, maths or multi-step planning. It does nothing for creative writing.
4. Set a hard limit
Without a limit you get padding. “In 150 words” or “in exactly 5 bullet points” produces tighter output than any instruction to “be concise”, which the model reliably ignores.
5. Tell it what to leave out
Negative constraints work better than people expect:
- No introduction, start with the first point
- Do not use the words “delve”, “leverage” or “landscape”
- No summary at the end
That last one alone removes most of the tell-tale AI cadence.
6. Make it ask you questions first
Before you answer, ask me up to 5 questions you need answered to do this well.
This is the pattern I use most and recommend most. It converts a guessing exercise into a conversation, and the questions themselves often reveal what you had not thought about.
7. Ask for three versions, not one
One output invites you to accept it. Three invites you to compare. Ask for versions that differ on a stated axis — one formal, one blunt, one warm — rather than three attempts at the same thing.
8. Give it the failure mode
“Here is what a bad answer looks like, and why” is unreasonably effective. Models are better at avoiding a described failure than at hitting a described success.
9. Separate instructions from content
When you paste a long document, mark it clearly:
Instructions: summarise the risks below in 5 bullets.
---
[document]
---
Without a separator, instructions inside the pasted text can get followed as if they were yours. That is a real problem with anything you did not write.
10. Ask it to critique its own output
Now list the three weakest points in what you just wrote.
The critique is usually accurate and the rewrite that follows is usually better. It costs one extra message.
11. Anchor to a real reference
“Write in the style of a good technical blog” is vague. Pasting three paragraphs of writing you admire and saying “match this register” is specific.
12. Stop starting over
The most common mistake is opening a new chat when the answer is wrong. The context you already built is worth more than a fresh start. Say what is wrong and ask for a revision.
A quick comparison
| Instead of | Try |
|---|---|
| Write a blog post about X | Ask me 5 questions, then draft 800 words for [audience] |
| Make it better | Cut 30% without losing the examples |
| Be concise | Maximum 120 words |
| Act as an expert | You are advising someone who will act on this and cannot afford to be wrong |
Common questions
Do longer prompts always work better?
No. Longer prompts work better up to the point where they stop adding information. A 300-word prompt full of constraints beats a 30-word one. A 300-word prompt that repeats itself does not.
Does saying “please” change anything?
Not measurably. Politeness costs nothing, but it is not a technique.
Do these work in Claude and Gemini too?
Yes. Every pattern here is about supplying information and constraints, which is model-agnostic. The specific wording matters far less than people think.
