Home · Insights · Implementation

How we run a business with Claude: the folder, the rules and the retrieval table

By the Viacala team · 9 October 2026 · 6 min read

The assistant is only as good as the context it starts with. Write the business down once, and every task starts from the truth instead of a guess. Here is the structure we use to run our own company.

Two different things called "projects"

Claude has a feature called Projects in its apps: a workspace where you upload documents to a knowledge base and set project instructions that apply to chats inside it. The second thing is what we do ourselves, which is run the company out of a folder of plain text files that Claude reads. They share an idea: the assistant is only as good as the context it starts with. Here is how we use that idea.

Why context beats prompting

Most people use AI by typing a fresh request each time and re-explaining who they are. That puts the burden on the prompt. A better approach is to write down, once, the things that are always true about the business, and let every task start from them. Your services, your prices, your tone, your ideal client, what you will never say. Put them in a file. Every output improves, and so does every output you did not write yourself.

The structure we use

• One instruction file at the top: who the company is, the working rules, and a retrieval table.
• A folder for what is always true. Business description, ideal client, offers, brand voice. Changing a file here changes every task that uses it, which is deliberate. We never copy that information into individual tasks.
• One folder per client, read in full before any work on that client.
• A folder of playbooks for jobs we repeat, so a job is done the same way each time.
• A dated log of what was done, so the next session starts knowing where the last one ended.

The retrieval table is the part that matters

A folder full of files is not useful if the assistant reads all of it or none of it. The retrieval table is a short list at the top: for client work, read the client folder and the offers file; for content, read the brand file; for strategy, read the business overview. It stops the assistant loading everything, which wastes its attention, and stops it missing the file that matters. Write yours as a table of task, then files. It takes ten minutes.

What it does not solve

It does not make the assistant correct. It reduces how often it is wrong about your business, because it starts from what you wrote and not from what it guessed. It does not remove the need to check anything that carries a price, a date or a promise, as we cover in what AI cannot do yet. And it does not work if the files go stale. Treat them like any other thing the company depends on: someone owns them and updates them when a decision changes.

Starting small

1. Write one page on what the business does, for whom, and how you speak.
2. Write the ideal client in five lines.
3. Add a retrieval table with three rows.
4. Use it for a week and fix what it gets wrong in the files, not in the chat.

We set this up for ourselves first, and it is how we organise the work described on how we organise AI work for Mallorca clients. The tools we use around it are in the AI stack we run Viacala on.