A copy-paste prompt that flips the direction of the conversation: your AI asks you ten questions, one at a time, pushes back on vague answers, then builds the plan only from what you said. Runs in the AI assistant you already use.
You ask your AI for a draft. It comes back generic: correct, polished, and belonging to no one. The instinct is to blame the model. The model isn't the problem.
Think about what a prompt actually is. A prompt is you guessing what the AI needs to know, then writing that guess from a blank page. Writing instructions about your own business is hard. And half the time you don't fully know what you want until something forces you to say it out loud.
Here's the part most people miss: this stays true even when the AI already knows you. Mine is connected to my files, my past conversations, my whole history, and it still produces generic work when I just throw a one-line prompt at it. Knowing everything about you is not the same as knowing what you want right now, for this specific thing. Your files are your past. Your intent lives in your head, and there is exactly one way for the AI to get it out: ask.
The fix is one move: before the AI produces anything, it interviews you. Questions flow toward you. Answers flow back. The plan gets built from your answers, and nothing else.
How to read it: two rows, left to right. Top row is the old way: you guess, the AI guesses, and the gaps get filled with generic filler. Bottom row is the flip: the AI asks first, you answer in your own words, it plays your answers back for correction, and only then does it build. The output changes because the input changed.
An AI that politely asks "any preferences?" is not an interview. These four rules are built into the prompt below, and each one exists because the interview fails without it.
Ten questions in one block gets you ten shallow answers. One at a time forces a real conversation, and each answer shapes the next question.
When you say "I want it to feel premium," the AI asks for the last time you saw premium done right. Stories carry the detail that adjectives hide.
The AI summarizes what you said and asks what it got wrong. Two minutes of correction here saves a full rewrite later.
Nothing invented, nothing imported from "best practices" you didn't ask for. If the AI would need to guess, it asks instead. That's what keeps the output sounding like you.
This is the full prompt. Copy the whole block, paste it into your AI assistant, and describe your task in the first line.
I want your help with something, but do NOT start working on it yet.
THE TASK. Here is what I want help with, in one rough sentence: [describe it — a plan, an email, an offer, a hire, a decision, anything]
STEP 1 — INTERVIEW ME FIRST.
Before you plan or produce anything, interview me. Ask me 10 questions, ONE at a time. Wait for my answer before asking the next. Your questions should dig out:
- what I actually want from this, and what "great" looks like to me
- who this is for, and what I know about them that you don't
- what I've tried before, what worked, and what I would never do
- my constraints: time, money, tone, non-negotiables
STEP 2 — PUSH BACK.
If my answer is vague, do not accept it. Ask for the story or example behind it ("tell me about the last time that happened"). If two of my answers contradict each other, point it out and make me pick.
STEP 3 — PLAY IT BACK.
After the ten questions, summarize what I told you in my own words and ask me what you got wrong. Fix it before moving on.
STEP 4 — BUILD.
Only now, produce the plan or the draft. Build it ONLY from my answers and any files I gave you. If something is not from my answers, leave it out. Wherever you would need to guess, ask me instead.
RULES.
One question at a time, never a list. Plain language, no jargon. My answers outrank your best practices.
The interview runs as a conversation. Block 15 to 30 minutes and answer out loud if that's easier: voice notes work.
ChatGPT, the Claude app, or the Codex app. Whichever you already use. If it already knows your business from earlier chats or files, even better.
Copy the whole block above, fill in the one-sentence task line, and send it as one message. The AI asks its first question.
Answer the ten questions honestly, fix the play-back summary, and take the plan. Everything in it traces back to something you said.