Best AI Prompts: 40 Examples That Work in ChatGPT, Claude and Gemini

How the forty were picked

The library has hundreds of prompts, and plenty of them do the same job in slightly different words. For this page we kept one per job and judged each candidate the same way. A prompt made the list when it did all of these:

1. It names its inputs. You can see at a glance what to paste or fill in, so you don't discover halfway through that it needed your budget or your deadline.
2. It fixes the shape of the answer. Numbered parts, in order, with lengths. Two runs of one prompt come back comparable, and you know where to look.
3. It forbids inventing. No made-up figures, names or dates; gaps get marked. The A/B Test Analyzer is built around saying what your numbers can't support.
4. It has a way out when your input is thin. Usually a line telling the model to ask a few questions before it starts.
5. It needs no feature only one assistant has. No browsing, no image editing, no app connection. That's what makes all forty portable.
6. It needed the fewest edits. Where two prompts covered one job, the one you can use with less rework won.

Five picks are older and shorter than the rest: the Commit Message Generator, Regex Generator, English Translator and Improver, Microsoft Excel Expert and Midjourney Prompt Generator. Their jobs are narrow enough that a short prompt does them, but you'll give them more detail yourself. The worked example below shows how.

Put the criteria together and you get a skeleton most picks share. You can write your own prompts on it:

Act as [the expert who does this job].
I will give you [input 1], [input 2] and [input 3].
If [the essential input] is missing, ask me up to three questions first.
Answer in [number] parts, in this order: [part 1], [part 2], [part 3].
Use only what I give you. Write [MISSING] instead of guessing.

Which pick for which job

The list on this page isn't in reading order, so here it is by kind of work.

How to tell whether a prompt is any good

You'll find prompts everywhere, and most look fine until you run them. These checks take a minute and catch most weak ones before you waste a chat on them.

The first check is Anthropic's own advice: show the prompt to a colleague with no context, and if they'd be confused, the model will be too. The third echoes OpenAI's guidance, which prefers «use a 3 to 5 sentence paragraph» to «fairly short». The last one is yours to run: try the prompt on the messiest real input you have, not a tidy one.

Worked example: giving a short pick the rest

The Microsoft Excel Expert is one of the short classics. As written, it opens a conversation and asks what you need. That's fine for exploring, slow for a specific problem. Here's how we'd extend it for one real task. This is our own example, written to show the checks above in use:

[paste the Microsoft Excel Expert prompt]

My problem: column B has dates typed as text, like "03/04/2026",
and some are US order, some UK. I need real dates in column C.
I use Excel for Microsoft 365 on Windows.
Answer in three parts: the formula, how to tell which rows
it got wrong, and what to do with those rows by hand.
If you can't tell US from UK order for a row, say so; don't guess.

The additions map to the table: what to paste (the column and the format), measurable output (three named parts), a ban on guessing, and a hard case named up front (dates that could be read either way). Use the same extension on the Regex Generator or the Commit Message Generator: one real example of the input, the result you expect, and the edge case you're worried about.

Moving a prompt between ChatGPT, Claude and Gemini

All forty run as written in any of the three. What changes is how you hand over your material and how you read the answer.

• Long material goes first, the ask goes last. Anthropic's and Google's prompting guides both say so for long inputs: put the document above your question. With the prompt block at the top, paste your document after it and end with one line repeating the task, such as «Using the document above, do parts 1 to 4.»
• Wrap pasted material in tags. The Long Document Summarizer asks for the text between <document> tags. Tags like that work in all three and keep your instructions apart from the text they act on. Google's guide suggests XML-style tags or headings; pick one and stay consistent.
• Add one example when the format matters. Google recommends always including examples. One sample line of what you want back fixes format drift faster than another paragraph of rules.
• Expect different voices. The same prompt comes back longer from one assistant and blunter from another. If length matters, the prompt should give a number; many picks already do.
• Know how each one takes files. In ChatGPT, as of October 2026, a paste over 10,000 characters turns into an attachment automatically. That's fine for these prompts, since the instructions stay in your message.

When you use one assistant only, its own page does more: ChatGPT prompts use uploads and memory, Claude prompts lean on long pasted documents, and Gemini prompts work with your Google files. Our ChatGPT vs Claude vs Gemini comparison covers which to use for what.

Follow-up lines when a pick misses

Good prompts still produce weak answers sometimes. These lines work the same way in all three assistants.

Where even the best prompts fall short

A prompt can't fix the input it's given, and it can't check the world for you. Code and formulas from the Regex Generator, the SQL Query Optimizer and the Excel picks have to be run on your own data before you trust them. The A/B Test Analyzer is only as good as the counts you paste. The Salary Negotiation Script and the Keyword Research Strategist won't invent salary data or search volumes, and they tell you where to look instead. Take that seriously and get the numbers from a proper source. And a prompt picked for being portable is, by design, a little less sharp in any single assistant than one written for it. That trade is the point of this page.