Most businesses try AI the same way. Open a chat, explain the company, ask a question, get a reasonable but generic answer. Next week, explain the company again. The model is not the problem: it knows nothing about you when the conversation starts, and forgets it when the conversation ends.
The fix we use, for our own company and for clients, is not another platform. It is a folder: plain text files that describe how the business works, kept under version control, which any AI tool reads before it does anything. Here is how it is set up, with a real example from Charissara, a fashion brand.
A real example: a fashion brand, from factory to online store
Charissara is a Thai womenswear label founded by Natthika. It designs its own patterns, sources fabric, produces in its own workshop and sells online and through social media. Before the workspace, what the business knew lived in chat threads, spreadsheets and Natthika's head.
Now it lives in one folder, with a subfolder for each part of the business:
- Brand: identity, tone of voice and the approved brand kit.
- Collections and products: every pattern, every fabric, size charts and pricing.
- Factory, quality and supply chain: how an order becomes a garment, inspections, suppliers and materials.
- Sales, website, marketing and customer service: channels, store briefs, campaigns and service standards.
- People, finance and legal: roles, budgets and contracts.
- Templates: the standard format for any new document.
What it changed
With that folder in place, the AI could do real work instead of giving advice. In one session it read the pattern and fabric libraries and created 132 product drafts in the online store, one per pattern and fabric pair, with 1,018 size variations, each priced from the pricing files and labelled with where its price came from.
It then wrote a photo brief for every product: which fabric photo and which shape reference to give an image generator, the prompt, and the file name, so each finished picture is matched back to its product automatically. The brand kit went onto the site through the theme's own settings, so Natthika can still change it herself from the admin.
Day to day, she now asks for what she needs in plain language, a price change across a fabric range, a new product from a pattern record, an update to the catalogue, and the AI works from the same files she does. Nothing goes live on its own: products are created as drafts, and she checks size, price, availability and photos before publishing.
Why plain files beat a platform
- Any AI can read them. Claude, ChatGPT, Gemini or whatever comes next: the folder does not care which model you use this year.
- You own them. No subscription holds your company knowledge hostage, and nothing is lost if a tool shuts down or changes its price.
- People read them too. A new hire reads the same pages the AI reads, so the documentation stays accurate because it is actually used.
- They cost nothing. A text file is free to store, search, copy and back up.
How to structure the folder
Structure matters more than volume. An AI handed a hundred unsorted documents does about as well as a new hire handed the same pile on their first day.
- One instruction file at the top: what the business does, who decides what, the rules that apply everywhere, and where things are. Keep it short, since it is read before every task.
- One folder per part of the business, each with its own short instruction file for the rules that only apply there.
- One source of truth. Link to a file rather than copying its content, so a price or a size chart lives in exactly one place.
- A history log: one dated entry per meaningful change, so anyone, human or AI, can see what happened and why.
- Decisions, owners and next dates at the top of every working document.
Let the AI act through simple tools
Reading is only half of it. The value comes when the AI can act: update a product, build a report, rename a batch of files. There are two common ways to give it that reach.
Connectors, often built on the Model Context Protocol (MCP), plug an AI app into a service such as your email or your calendar. They are convenient for looking things up and for occasional actions.
For repeated operational work we prefer command-line tools: small programs the AI runs the way a person would type them. Many systems already ship one, WordPress has WP-CLI, and a missing one is usually a short script. They are easy to test, every action leaves a line you can read afterwards, they work with any AI that can run commands, and they do not fill the model's working memory with descriptions of tools it will never use in that task.
Version control as memory and safety net
Put the folder in Git, in a private repository on GitHub or a similar host. Every change is recorded with who made it and when, any mistake can be undone, and the same folder stays in sync across every computer and every person working on it. A scheduled sync means nobody has to remember to save.
It also changes how you review AI work. Instead of trusting a summary of what was done, you read exactly what changed, line by line, before keeping it.
Guardrails before anything else
- No passwords or API keys in the files, ever. They stay in the systems they belong to.
- Drafts first. The AI prepares, a person publishes.
- Live systems are read only unless someone explicitly asks for a change.
- Rules are written down. When the AI gets something wrong, the fix is a new line in the instruction file, not a hope that it remembers.
How to start this week
- Create one folder and write one page: what the business does, who the customers are, and the five rules everyone should know.
- Add a subfolder for the process that costs you the most time, and describe it as the steps people actually take.
- Put the folder in a private Git repository.
- Open an AI tool that can work with files, such as Claude Code or Codex, inside that folder, and ask it for one real task from that process.
- Each time it gets something wrong, add the rule. After a few weeks, the folder is the most accurate description of your business you have ever had.
Where we come in
Setting up the folder is a few days of work, and most of it is getting what sits in people's heads onto the page. The part that takes experience is the connection: giving the AI safe, tested tools for your store, your CRM or your internal systems, with drafts, logs and limits in the right places.
That is the work we do for clients, and the same setup we use to run our own company.