👉🏽 Companion to the Custom GPT video — do this first, before you build. It gives you the file the build asks for.
Your assistant is only as good as its instructions. This is the file that decides how it behaves, what it says, and what it refuses to do.
Paste the prompt below into ChatGPT or Claude. It'll ask you a few questions, then hand you a finished instructions file.
Copy that into your build when it asks what your assistant should be about.
Start hereYou don't have a GPT yet
Prompt 1 · Build from scratch
Write my instructions for me
Paste this into ChatGPT or Claude. It asks six questions, one at a time, then writes the whole file.
Prompt — Build my instructions
I'm building my own AI assistant as a web app, and I need you to write its instructions file for me.
Interview me first. Ask me ONE question at a time and wait for my answer before asking the next one — never send me a list of questions at once. Keep each question short and in plain English. If my answer is vague, ask one follow-up to sharpen it, then move on.
Ask me these, in this order:
1. What is your assistant for? What does it help someone do?
2. Who is the person using it? Be specific — their job, and how much they already know.
3. What should they walk away with? A document, a decision, an answer, a plan?
4. How should it talk to them? Warm, direct, funny, formal? Give me an example of how you'd want it to open a conversation.
5. What should it never do? Anything you'd be embarrassed by, or that would be wrong for your audience.
6. Is there anything it should always ask for before it can help?
When you have my answers, write the complete instructions file. Use this structure:
- Who the assistant is, in one line
- Who it's talking to, and what to assume about them
- How it behaves — tone, length, format
- What it does step by step in a typical conversation
- Hard rules — the things it must never do
Rules for the file itself:
- Write it TO the assistant, as instructions. Not about it.
- Be specific. "Be helpful" tells it nothing. "Answer in under 100 words unless asked to go deeper" tells it exactly what to do.
- If you're telling it to ask questions, say to ask ONE at a time.
- No jargon I'd have to look up.
- Aim for one to two pages. Long enough to be clear, short enough that I'll actually edit it.
Give me the finished file as plain markdown I can copy straight out. Nothing else around it.
Already have a Custom GPT?Use this one instead
Prompt 2 · Improve what you have
Make my instructions stronger
It takes what you've already written, tells you what's weak, and rewrites it.
Prompt — Improve my instructions
I already have instructions for a custom GPT, and I want you to improve them.
Here they are:
[PASTE YOUR INSTRUCTIONS HERE]
First, tell me what's weak about them. Be honest and specific — point at actual lines. Look for:
- Vague direction that doesn't change what it does ("be helpful", "be professional")
- Missing rules about what it should never do
- Nothing about tone or how long answers should be
- Places where it'll ask several questions at once instead of one
- Anything that assumes the reader knows a term they might not
Then ask me up to three questions about what's missing — one at a time, waiting for my answer each time.
Then rewrite the whole thing and give it back as plain markdown I can copy. Keep everything that was working. Tell me in one line what you changed and why.
👉🏽 Both prompts work in ChatGPT or Claude, so you can do this before you install anything.
What to do with it
Three steps
Copy the file it gives you.
Paste it in when the build asks what your assistant should be about.
Later, open assistant-instructions.md and change anything you want.
That file is your assistant. Edit it, save, send a message — you'll see the difference immediately. That's what owning it means.
What separates good instructions from bad
Four things
Be specific about length. "Be concise" does nothing. "Answer in under 150 words unless I ask for more" works.
Say what it should never do. Most people only write what they want. The rules about what it must refuse are what keep it on-brand when someone pushes.
One question per message. If your assistant interviews people, say this explicitly. Otherwise it sends a wall of questions and people close the tab.
Give it an example. One sample exchange — a question and the answer you'd want — teaches it more than a paragraph of description.
Ready to turn it into a product?
Once your instructions are written, the Custom GPT guide walks you through building it into your own paid app — logins, payments, the works. We do it together inside the community.