The 15-minute setup that turns ChatGPT or Claude from a clever search box into something that knows you, your work, and how you like things done.
You're already using ChatGPT or Claude. This is the setup that turns it from a clever search box into something that actually knows you, your work, and how you like things done. No code required.
It interviews you for about five minutes, then writes your reusable AI setup files. This is the whole guide compressed into one prompt. Everything below explains what it's doing and where the files go.
You're going to help me set up a reusable context profile for working with AI tools. I want files I can reuse in every session so I don't re-explain myself every time. Interview me one question at a time. Do not ask everything at once. Cover: 1. My work — my role, what I'm responsible for, what I'm trying to get done over the next few months. 2. My tools — what I work in day to day (apps, file types, whatever's relevant to me). 3. How I like to work with AI — plan first or just do it, how blunt you should be, response length, anything that annoys me. 4. My recurring tasks — the things I'd hand to an AI every week if it already knew my context. 5. If I create content — who it's for, my tone, one or two things I'd say and one or two I never would. Skip this if it doesn't apply to me. Ask a follow-up when an answer is vague. When you have enough, output three files as separate Markdown code blocks: - about-me.md — who I am and what I'm working on - working-preferences.md — how I want you to collaborate with me - brand-voice.md — only if I create content Keep each file under ~200 words, in my words, no generic filler. Ask me the first question now.
Most people only ever use the first one. That's like buying a Swiss Army knife and only using the toothpick. Here's the whole knife, and you don't need to be technical to use any of it.
You type, it answers. Great for drafting, summarizing, and questions. It forgets everything between sessions, and that wall is where most people quietly give up on AI.
It reads your files, runs commands, and does multi-step work in your real environment. Same conversational feel, except now it can actually touch your stuff.
Describe a task, walk away, come back to a finished result. Works best on clear jobs you could hand to a competent junior with a short brief.
This is the part that separates casual users from people who get real work done. The same model gives two people wildly different results, and the difference is almost always this: one of them set up context once, the other re-explains themselves every session.
Open Claude or ChatGPT, paste the prompt from the top of this guide, and answer its questions like you're briefing a sharp new contractor on day one.
# About Me I run marketing and ops for a 12-person SaaS company. I'm not an engineer. I live in Notion, Gmail, Google Sheets, and our analytics dashboard. I write in plain English and I'm comfortable following step-by-step instructions, including the occasional terminal command if you walk me through it. ## What I'm working on - Q3 launch: landing page copy, email sequence, launch checklist - Monthly metrics report I rebuild by hand every time (hate this) - Keeping our help docs current as the product changes ## What I'm not A developer. Skip the jargon. If a task needs a tool I don't have open, tell me what to install and why before we start.
Same files, different homes depending on your tool. Find the row that matches how you work and follow it.
Claude.ai → Create a Project, paste files into Project knowledge ChatGPT → Settings → Personalization → Custom Instructions Claude Code → Save in your working folder as CLAUDE.md Cursor → Save as .cursorrules in your project root
# How I Like To Work
- Show me a short plan before anything multi-step. I'll say go.
- Be blunt. If my idea is weak, say so and tell me why.
- Keep answers tight unless I ask you to go deep.
- Don't produce final copy. Give me two or three options.
- When you're unsure, ask one sharp question instead of guessing.
You don't need a new tool. You need the one you have to stop forgetting who you are. This is five minutes and it changes every conversation after it.
ChatGPT calls it Custom Instructions. Claude calls it Projects. Same idea: context that survives between chats so you stop repeating yourself.
Here's context about me. Use it in every reply unless I say otherwise. WHO I AM: [your role, what you're responsible for] MY TOOLS: [the apps and files you work in] HOW TO HELP ME: be direct, plan before multi-step work, keep answers tight, give me options instead of one final version, ask a sharp question when you're unsure instead of guessing. Acknowledge in one line, then wait for my first real task.
What should ChatGPT know about you? Marketing/ops lead at a small SaaS company. Not technical. Work in Notion, Gmail, Sheets. Currently running a Q3 launch. How should ChatGPT respond? Be direct and brief. Plan before multi-step tasks. Give 2-3 options, not one polished answer. Plain English, no jargon. Ask one question if unsure instead of guessing.
Not "write me a poem." Real work builds intuition that toy prompts never will. Pick three things you actually have to do.
1. Paste a long doc you need to read → "Summarize this for someone in my role. What needs a decision from me?" 2. The email you've been avoiding → "Draft a reply. Match the tone of my last message in this thread." 3. Something you're learning → "Explain this like I have 15 minutes and a real reason to care, not a textbook."
Once you move past chat, every serious tool reads a plain text file at the start of each session to learn your project. The names differ. The idea is identical: a file the AI reads so it doesn't guess.
| Tool | File name | Where it lives |
|---|---|---|
| Claude Code | CLAUDE.md | Root of your project |
| OpenAI Codex | AGENTS.md | Root of your project |
| Cursor | .cursorrules | Root of your project |
| GitHub Copilot | .github/copilot-instructions.md | The .github/ folder |
| Emerging standard | AGENTS.md | Root of your project |
Ten lines. This alone stops the most common failure: the AI producing something that technically works but ignores how you do things.
## Project Overview This is a Next.js store for selling custom furniture. Stack: TypeScript, Tailwind, Supabase. ## Conventions - Named exports, not default exports - Tests live next to the file they test - Run tests with: npm test
The terminal sounds intimidating. The interaction is the same as chat: you type plain English, it responds. The difference is it can read and change real files on your computer. You only have to do this once to get it.
One command. If you don't have Node or Homebrew, ask ChatGPT "how do I install Homebrew on a Mac" first and follow it.
npm install -g @anthropic-ai/claude-code # or, if you use Homebrew: brew install claude-code
Make a folder of files you actually have (notes, a messy spreadsheet export, a pile of docs). Open it, run claude, and try this.
Look at every file in this folder. Tell me what's here in plain English, flag anything that looks duplicated or out of date, and propose a cleaner folder structure. Show me the plan before you move or rename anything.
Cloud agents work while you do something else. The skill isn't technical. It's learning to scope a task tightly enough that a competent stranger could do it from your description alone.
This single distinction is most of the value. Vague task in, mediocre result out. Every time.
✗ Bad: "Build me an app." ✗ Bad: "Improve the signup flow." ✓ Good: "Add email-format validation to the signup form. Show an inline error. Write a test for it." ✓ Good: "Add a dark-mode toggle to the settings page. Persist the choice. Match existing styles."
Paste it into Codex or push it as a background task in Claude Code. Then close the tab and come back later.
TASK: [one specific, self-contained change] DONE WHEN: [how I'll know it worked — the visible result] CONSTRAINTS: [match existing patterns, don't touch X, etc.] IF UNSURE: stop and leave a note rather than guessing. Show the change as a pull request I can review.
Every tool here forgets between sessions unless you give it somewhere to remember. The fix is not a product. It's a folder of plain text files the AI can read and write directly.
A folder of markdown notes: decisions, preferences, project context. Point your AI at it. Not sophisticated. Works today, costs nothing.
Fine for you, awkward for AI. The content sits in a format agents can't read directly, so you're back to copy-pasting. Friction returns.
Everything is plain markdown in a local folder. No proprietary format, no export step. The agent reads and writes it natively. This is what I use.
You don't have to do all of this. Pick the column that sounds like you and do those three things this week.