Stop Writing Prompts. Start Writing Instruction Files.

Halfway through an AI lunch-and-learn for a product company with a few dozen people, a hand goes up.

“When I start a new chat, do I have to tell it to remember things? Because sometimes it remembers and sometimes it doesn’t.”

It’s the most honest question of the day. Everyone who uses AI at work pays the same tax: explaining themselves, again, to a machine that seems to forget them overnight. Except on the days it doesn’t.

I ask how they use it. In the browser, a fresh chat each time.

So that’s the answer, and it isn’t the one anyone wants. It never remembered anything. Every new chat is a blank window. Nothing is loaded. When it does seem to remember, the app is deciding on its own which scraps of past conversations to carry forward, and it doesn’t keep everything. So sometimes it remembers. And sometimes it doesn’t.

I know that blank window well. I used to fight it too.

I was using AI to help me evaluate new business opportunities, and it kept giving me the wrong advice. Not because the model was weak. Because it had no idea who was asking. It didn’t know I run Seller Labs, or what Seller Labs does, or who we sell to. So it gave the advice you would give anyone, which is advice for no one.

Then I started letting it fill out forms for me, and it kept getting my details wrong. Including which of my email addresses goes where.

So I told it who I am, and I asked it to write that down.

The next new session, I don’t introduce myself.

It already knows.

That moment delighted me more than any answer it had ever given me, because I could see where it went. If it could remember who I am, it could remember how I work. What I like. What I don’t. Everything I teach it once, it keeps.

What made that work was not a clever prompt, and not a memory setting. It was a text file. Here are a few lines from the first version:

## Who I Am
- Nhan Dinh. CEO & Product Manager at Seller Labs
- I make product and technical decisions. Respond at a business
  + technical level (not overly simplified, not overly academic).

## Company & Product
- Seller Labs: B2B SaaS platform for Amazon marketplace sellers

## Response Preferences
- Be concise and direct. I value efficiency
- Don't over-engineer. Keep solutions simple and focused on what was asked

That’s it. No code. My agent reads it at the start of every session, so every session starts already knowing me.

It grew the way a good employee’s notes grow. Every time the agent got something wrong, I added a line. Which email goes on vendor forms and which one is for work. How I like files named. What it must never do without asking me first. The page that started it all now has a 157-line main file, 71 memory files, and more than a hundred project folders, each with its own context file.

It isn’t magic. When a rule gets buried too deep, the agent can still miss it in the middle of a task. So the rules that matter most live at the top of the main file, where it can’t skip past them.

Back in the room, the team’s own question 7 comes up later: “What are some good prompts for streamlined emails?”

It’s the same question in a different outfit.

“That question means you’re still using the chat prompt,” I tell them. The value moved. It’s not in writing prompts anymore. It’s in writing instruction files.

Write it down once

Start with who you are. An instruction file is usually a markdown file, which is just a plain text file the AI reads. Each tool looks for its own file name (Claude Code reads CLAUDE.md, OpenAI’s Codex reads AGENTS.md, Gemini CLI reads GEMINI.md), but what goes inside is the same: your name, your role, what your company does, who you serve, and how you like answers. Write it the way you’d brief a sharp new hire on day one.

Give each project its own file. Mine is one folder per project, each with a short context file: what the client does, what we’re building, what’s already decided. When I work in that folder, the agent loads that file automatically. I never re-explain a project.

Teach it your voice from your sent mail. AI is a great mimicking machine. Point it at your past emails, ask it to describe how you write, and save that description in the file. That’s also why I don’t reach for the AI button built into an email or ERP tool to draft a reply. It doesn’t have my context. My agent does, and it compounds.

Add a line every time it gets something wrong. Every correction you type in a chat evaporates when the window closes. Every correction you write in the file stays fixed for good.

Keep the rules that matter at the top. The file gets long. Put the non-negotiables first.

If you want the step up from chat to an agent that can open these files, Article 3 in this series walks through it.

Still working in a chat window?

Projects are the halfway step. Both ChatGPT and Claude let you set instructions and upload files that apply to every chat inside a project, so start there.

And when you do write a one-off email prompt, use four parts:

1. Who you are and who they are. “I’m a sales rep at a gear company, writing to a dealer who has ordered twice.”
2. What you want to happen. The outcome, not the topic. “I want them to agree to a fifteen-minute call.”
3. The actual context. Paste the thread. Almost nobody does this, and it’s the biggest quality difference.
4. Constraints. “Under a hundred words, no exclamation points, sound like a person.”

Three patterns worth keeping:

• Reply drafting: paste the thread, say what you want to happen, get a draft.
• Tone repair: “Make this firm but not rude.”
• Extraction: “What did I commit to in this thread, and by when?”

Then notice which parts you keep typing. The parts you type every time belong in the file.

So, do you have to tell it to remember?

Yes. Once. In a file it reads every time.

Then stop introducing yourself.