Folder Brain

Build your folder brain

Your AI is not confused. Your business is still mostly in your head. Write it down once, in four plain files you own, and any AI can pick it up. Including the ones you have not tried yet.

Four plain filesOne folder you ownWorks with any AI

Sources opened and quoted on 16 September 2026

The problem, in one line

You open a new chat. You explain your business. You get an answer. You close the tab. Tomorrow you explain it again.

Everyone is busy building AI agents and giving them names. Almost nobody has written down what the business is. An agent with no context is a new hire on day one with no briefing.

So write it down once. Four files, one folder, on your own computer.

The four files

Four questions. Make a folder, put four files in it, and answer one question per file. Nothing needs to be long. A page each is plenty to start.

Agents go underneath, not beside. An agent is only as good as the context it stands on.

Brief

01-brief.md

What is this business, and who is it for?

What you sell, who you sell it to, what you are trying to hit this quarter, and how you sound. Add the words you use that an outsider would not understand.

Engineers call this domain context.

Boundaries

02-boundaries.md

What is the AI allowed to do?

What it may do on its own, what it must never do, and what it has to bring back to you first. Write drafting and sending as two different things. Same with preparing and spending.

Security people call this least privilege.

Breadcrumbs

03-breadcrumbs.md

Where does the current version live?

Where the live version of each thing sits. Your client list, your prices, your calendar, your files. Point at it. Do not paste a copy in, because then you have two and one of them is wrong.

In IT this is your system of record.

Blueprints

04-blueprints.md

How does a job get done here?

The steps for one job you repeat. Write the order, and write how you know it is finished. Be blunt about it: a good cold email for my software does name the problem in the first line and never opens with a compliment. That sentence is worth more than a page of description, because the AI can check its own work against it.

In operations this is a runbook, and the finished test is your acceptance criteria.

Do not start from a blank page

Staring at four empty files is how this stalls. So do not write them. Get interviewed instead.

Paste this into any AI chat. Answer the questions out loud if it is easier, then paste your answers in. At the end it hands you the four files.

The interview prompt

I want to write down the context of my business so any AI can work with it.

Interview me. Ask me one question at a time and wait for my answer.

Cover these four things, in this order:
1. Brief. What the business is, who it serves, what matters right now, and how I sound.
2. Boundaries. What AI may do on its own, what it must never do, and what it brings back to me.
3. Breadcrumbs. Where the current version of each important thing actually lives.
4. Blueprints. The steps for the jobs I repeat, and how I know each one is done.

Rules:
- One question at a time. Short questions.
- If my answer is vague, ask me for a concrete example before moving on.
- Do not write anything until you have asked about all four.
- At the end, give me four separate markdown files: 01-brief.md, 02-boundaries.md,
  03-breadcrumbs.md, 04-blueprints.md.
- Use my words. Do not tidy my voice into something more corporate.

Save what comes back into your folder. That is your first version. It will be wrong in places, and that is fine. You fix it the second time you have to correct the AI about the same thing.

Why plain files, and not an app

Four markdown files in a folder are not a feature of anything you pay for. Nobody can move them to a higher tier. Nobody can retire them. If you stop paying for a tool tomorrow, the folder is still on your computer.

That is the whole reason to keep it outside an app. Your context is the expensive part. The model is the part you should be able to swap.

And there is a mechanical reason, not just an ownership one

Every AI chat has a limit on how much it can hold at once. When a long conversation reaches that limit, it does not simply stop. It summarises itself and carries on from the summary. Anthropic calls this compaction and describes it like this:

“Compaction is the practice of taking a conversation nearing the context window limit, summarizing its contents, and reinitiating a new context window with the summary.”

Their own warning about it is the part that matters to you: “overly aggressive compaction can result in the loss of subtle but critical context”.

So everything you explained in the chat is at risk the moment the window fills. Everything in a file is not. That is not a preference, it is how the thing works.

Anthropic, Effective context engineering for AI agents, page dated 29 September 2025. They have a name for the fix too, and it is the same idea as this page: “Structured note-taking, or agentic memory, is a technique where the agent regularly writes notes persisted to memory outside of the context window.”

The catch, and nobody selling you a setup mentions it

The folder travels. The wiring does not. Every tool wants its own small file at the top of the folder telling it where to look. When you move tools, you rewrite that one file. Not your business, just the doorbell.

Claude and Claude CodeCLAUDE.md
Codex, and most coding toolsAGENTS.md
Gemini CLIGEMINI.md
A chat window with no file accessPaste the files in, or attach them

I have moved one folder between two tools and had to write that file by hand. That is one test, on one folder, so treat it as the shape of the problem rather than a measurement.

One key, most of the models: OpenRouter

Normally, trying a different model means another account, another card, another key. OpenRouter is the shortcut. In their words:

“OpenRouter gives you access to hundreds of AI models through a single API endpoint.”

OpenRouter documentation, quickstart.

One account, one key, and you can change which model answers by changing one line. Your folder does not change at all.

How to start

1Make an account at openrouter.ai and add a small amount of credit.
2Create one API key. That single key reaches every model in their catalogue.
3Open your account settings and turn off routing to providers that train on your data. Do this before you send anything real.
4In whatever tool you are using, set the model name, for example deepseek/deepseek-v4.1-flash, and paste your key.
5Point it at your folder. Same four files, different model.

Step 3 is the one people skip

Different providers handle your data differently. OpenRouter says: “If you opt out of training in your account settings, OpenRouter will not route to providers that train.” That setting is yours to turn on, and it is off until you do. It does not make anything private, it changes which providers you get routed to.

What the Chinese models are

You keep hearing names like DeepSeek, Qwen and Kimi. They are large AI models built by Chinese companies, and two things make them worth knowing about.

They cost less. Often much less, for work that is good enough for a lot of everyday jobs.

They publish their weights. That means the model itself is a file anyone can download. You can run it on a computer you control, and nothing you type leaves the building. The big American labs mostly do not do this with their best models.

Who	Model	In	Out	Downloadable
DeepSeekChina	DeepSeek V4.1 Flash	$0.15	$0.60	Yes
MiniMaxChina	MiniMax M3	$0.30	$1.20	Yes
Z.aiChina	GLM 5.3	$1.40	$4.40	Yes
AlibabaChina	Qwen3.8 Max	$2.00	$6.00	Yes
MoonshotChina	Kimi K3	$3.00	$15.00	Yes
AnthropicUS	Claude Opus 5	$5.00	$25.00	No
OpenAIUS	GPT-5.6 Terra	$2.00	$12.00	No

Price is US dollars per million tokens of text going in and coming out. A token is a chunk of text. Anthropic’s glossary puts it at “approximately 3.5 English characters”, and says that varies by language. Pulled from the OpenRouter catalogue on 16 September 2026, where 444 models were listed. Downloadable means the weights are published on Hugging Face, checked the same day. These prices change. Look them up before you rely on them.

Three ways to reach them

Through OpenRouter

Easiest. One key, and you type the model name, for example deepseek/deepseek-v4.1-flash. You are still sending your text over the internet to whoever is hosting it.

Direct from the company

Each lab runs its own website and its own API. More setup, one more account, and your data goes to that company under its own terms.

Downloaded, on your own machine

Because the weights are public, you can run one locally with a tool like Ollama or LM Studio. Nothing leaves your computer. You need a capable machine, and the smaller models that fit on a laptop are weaker than the hosted ones.

Think about this before you paste client information anywhere

Using a hosted model means sending your text to a company, under that company’s terms, often in another country. That is true of American providers too. If you handle health data, client records, or anything covered by a contract or by GDPR, read the terms of whoever is actually hosting the model, and check what you are allowed to send. Running a downloaded model on your own machine is the one route where the text does not leave.

What I have not tested

I have not run my own workspace on a Chinese model, or on a model running on my own machine. The files are plain text so they will be readable. Whether the answers are as good is a separate question, and I have not measured it.
The prices in the table were pulled from OpenRouter on 16 September 2026. They change. Check the live number before you budget anything.
Cheaper does not mean better for your job. A model that costs a tenth as much can still cost you more in time.

Why I care about this

I have a master’s in health informatics, the science of how information gets stored, moved, and understood by whoever picks it up next.

My first-author paper looked at why public health data was hard to share in one Canadian province, from ten interviews in 2019. The conclusion, in the paper’s own words: “the majority of barriers faced in this province are human-related”, and those barriers are “outside the scope of a technical solution”.

That is this, exactly. The tool is not the problem. Nobody wrote anything down.

Murphy S, Reilly P, Murphy T. Assessing the potential use of blockchain technology to improve the sharing of public health data in a western Canadian province. Health and Technology 2021;11(3):547–556. Read the full paper on Europe PMC, free.