The Claude Master Prompt: Make Claude Rewrite Your Prompts Before It Answers

You’ve seen the prompt formulas. Role, task, context, format, examples, tone. They work, and almost nobody remembers them at 9pm when they need an email written. Here’s the shortcut I use instead: I don’t write good prompts anymore. I write rough ones, and I make Claude turn them into good ones before it does anything. It’s one paste, it takes thirty seconds, and the result is better every single time, because Claude knows what a complete prompt looks like far better than any formula does. This guide is that paste, how to use it, three worked examples, and what to do when it misbehaves.

Not because it’s smarter than you. Because it knows exactly what it needs. Every weak answer you’ve ever gotten came from a detail that wasn’t in the prompt: who the thing was for, how long it should be, what to leave out, what tone. You didn’t include those because you didn’t think of them. Claude thinks of them every time, because they’re the things it reaches for first when it starts writing.

So instead of you guessing what to include, you flip it. You hand Claude the rough version and ask: if you were asking another AI to do this, how would you write the prompt? It writes the prompt it wishes it had gotten. It adds the audience, the length, the format. And where it doesn’t know something, the master prompt tells it to ask instead of guess. That last part is the whole trick. Guessing is where answers go wrong. Questions in brackets take the guessing out.

Paste this at the top of a new chat, put your rough request where it says, and send.

Before you do anything, rewrite my request as the prompt you would want to receive. Read what I typed below. Then write the version you would use if you were asking another AI to do this. Include: who the result is for, the goal and what done looks like, the length and format, the tone, what to leave out, and any facts you need from me. Where a detail is missing, put a question in brackets instead of guessing. Show me only the rewritten prompt. Do not do the task yet. Wait for me to answer the questions and say go.

My request: [paste it here]

The quick version from the video. When you don’t want the full paste, type your request as usual and add this one line at the end:

If you were asking another AI to do this, how would you write the prompt? Show me the version you'd use.

The full paste is better because of two things the short line doesn’t say: put questions in brackets instead of guessing, and don’t do the task yet. Use the short one when you’re in a hurry. Use the full one when it matters.

Make it automatic. In Claude, open a Project (they’re on the free plan too), click Set project instructions, and paste this version, which drops the “My request” line and tells Claude to treat every message as the request:

Before you do anything with any message I send, rewrite it as the prompt you would want to receive. Write the version you would use if you were asking another AI to do this. Include: who the result is for, the goal and what done looks like, the length and format, the tone, what to leave out, and any facts you need from me. Where a detail is missing, put a question in brackets instead of guessing. Show me only the rewritten prompt. Do not do the task yet. Wait for me to answer the questions and say go.

Every chat in that project now starts by rewriting whatever you type. In ChatGPT, paste the same text into Customize ChatGPT under custom instructions, or into a Project’s instructions. In Gemini, make a Gem with it. After that you just type rough requests.

Claude hands back a prompt with a few [bracketed questions] in it. This is the part people overthink. Don’t.

• One line per bracket. “For my landlord.” “Short, five sentences.” “Firm but polite.” That’s plenty.
• “Skip” is an answer. If a bracket asks for something you don’t care about, type skip and Claude uses a sensible default.
• Answer in a list, not by editing the prompt. Type “1. my landlord 2. five sentences 3. firm, polite” and then the word go. Claude fills in its own prompt and runs it.
• Say go. That’s the trigger the master prompt is waiting for. Until you say it, Claude keeps holding.

The whole exchange is: your rough request, Claude’s prompt with three or four brackets, your one-line answers, go, the result. Under a minute, and the result reads like you thought about it for ten.

Keep what it writes. The prompt Claude hands back is a finished, reusable prompt. Copy it into a note before you say go. The third time you ask for the same kind of thing (a client email, a caption, a weekly plan), you already have a prompt with brackets in it, and you’re only filling in the blanks. That’s a prompt library that builds itself from your own requests, in your own words.

Each one is the rough request you’d actually type, then the kind of prompt Claude hands back. The brackets are Claude’s questions, and they’re the point.

1. The email you’ve been putting off

Rough request:

write an email to my landlord about the leak under the sink, it's been two weeks

What comes back looks like this:

You answer: “1. Mr. Alvarez 2. August 30 3. by text 4. yes, the cabinet floor is soft now. go.” And you get the email you’d have spent twenty minutes writing.

2. The post that has to sound like you

Rough request:

write an instagram caption about my new fall grooming special

What comes back looks like this:

Notice the last bracket. It asked for a sample of your voice. You’d never think to include that. Claude always does.

3. The thing you don’t know how to ask for

Rough request:

help me figure out if I should switch my daughter's school

What comes back looks like this:

That’s a better conversation than “help me figure out” would ever have started. Same you, same Claude, better prompt.

It doesn’t know your facts. The brackets are Claude admitting that. The dates, the names, the price, the thing that happened: those come from you, and the answers are only as good as what you type into the brackets.

It doesn’t replace your judgment. Example three is a thinking partner on purpose. Claude can lay out a decision clearly. Making it is still yours.

It works best with a real task. “Make me money” is a wish. “Give me three ways to add $500 a month to my grooming business” is a task, and Claude turns a task into a prompt worth running. When you’re not sure what to ask for, use example three: ask Claude to be a thinking partner and let its questions find the task.