How to Build a Context File System for Your AI (Copy This Template)

A copy-paste starter template, a bootstrap prompt, and the exact steps to get Claude, ChatGPT, or Cowork to actually load your context files, so you stop re-explaining yourself every session.

You need two files, a prompt to build them, and about twenty minutes. That's the whole starting kit: a file with the facts and rules that never change, a file with what's true right now, and a way to get your AI tool to actually read both before you type a word.

I wrote about why I built mine and what it looks like at full size in How I Built My AI a Memory System. This post skips the story and gives you the reusable part: the template, the prompt, and the setup steps for the tools most people are actually using.

The two files you actually need

Start with two, not four. Everything else is an upgrade you add once you actually catch yourself repeating something, not a day-one requirement.

A rules file. Who you are, what you do, and the handful of things that never change: your tone, your pricing floor if you have one, words you never use, how you like work handed back to you.

A state file. What's actually true right now. Active projects, what's blocked, what's next. Nothing older than this week belongs in it. Once something's done or stale, it moves out.

That's the whole starting kit. A rules file and a state file solve the two most common failures: retyping the same background every session, and an AI confidently working off information that stopped being true.

Copy this starter template

Paste this into a plain text file, fill in the brackets, save it as rules.md and state.md (or about.md and status.md, the name doesn't matter).

# Rules

## Who I am
[2-3 sentences: what you do, who you serve, how you like to work]

## Hard rules
- [A tone or voice rule]
- [A word or phrase to never use]
- [A standing constraint, like a pricing floor or a client you don't take]

## Always ask before
- [Anything you never want done without your OK first]

# Current state

## [Project or client name]
Status: [one line]
Blocker: [one line, or "none"]
Next move: [one line]

Keep each project entry to three lines. If a status line starts growing past that, it's turning into a diary, and a diary is exactly what a state file isn't for.

The bootstrap prompt: let your AI build it for you

You don't have to fill the template out alone. Paste this into a new chat with whatever AI tool you're using:

This does two things at once. It gets you a working first draft in one conversation instead of staring at a blank template, and it forces you to actually answer each question instead of skipping the ones that feel obvious.

Getting your AI tool to actually load it

A file sitting on your desktop does nothing on its own. It has to load into the conversation automatically, and how that works depends on the tool. If you're not sure which to use, start with Cowork, it's the one built for people who aren't developers.

Cowork. Start here if you're not sure which tool to use. Same CLAUDE.md file as the others, but the folder has to actually be connected first. For an ad-hoc chat, connect the folder when you start. For a Cowork Project, link it under that project's Context section, that's the step people skip. Once it's connected, CLAUDE.md inside it gets read automatically. This is the exact mechanism SpringOS runs on.

Claude Projects. Create a Project, then upload both files to the project's Knowledge section. For extra steering, paste the rules file into project instructions too. You'll find that option under “Set project instructions” in the project menu. Anything in Project Knowledge loads into every conversation you start inside that project. You don't re-upload it each time.

Claude Code. Save the rules file as CLAUDE.md in your working folder. It's read automatically at the start of every session in that folder, no separate setup step, same underlying mechanism as Cowork, just without a folder-connection step, since Claude Code already runs inside whatever folder you opened it in.

ChatGPT. Create a Project, then open Project settings and paste your rules file into the project instructions field. Upload your state file to the project's files. Keep both files in that same project rather than splitting the rules file into your account-wide Custom Instructions instead, project instructions override Custom Instructions the moment you're chatting inside a project, so splitting them means the rules file quietly stops applying right where you'd be using it most. ChatGPT also has a separate Memory feature that saves facts about you automatically, and you can edit what it's saved, but it's built for picking up general context over time, not for holding a rules file you want followed exactly.

Any other tool, or no persistent feature at all. Paste both files at the top of a new chat. It's still faster than retyping from memory, and it's the fallback every other method is really just automating.

What to do after day one

Once you catch yourself explaining the same thing twice in a week, that's the signal to add a third file, not before. A history file for what's already happened, so it stops piling up in your state file. A task-specific file for a job you do often, like a voice guide or a client checklist, that only gets loaded when that specific job comes up.

Three or four files still need upkeep, or they rot the same way one file would. How often depends on how often you actually use it. If you're in Claude daily, check in daily, and that's why mine runs daily now. If it's more of a once-a-week tool for you, weekly is plenty, since there's less piling up in between. Clear anything finished out of the state file and into the history file. If you've explained the same thing again, that's usually the sign a task-specific file is overdue. Automate this if your tool supports scheduled tasks. It's a five-minute pass, cheap insurance against exactly the staleness problem this setup exists to prevent.

Sources: Claude Help Center: How can I create and manage projects?, OpenAI Help Center: Projects in ChatGPT