How I run two businesses from one Claude Cowork workspace

Most people use AI like a search box. Fresh chat, ask a question, get an answer, close the tab. The next morning they open another blank chat and explain who they are all over again. The assistant never learns a thing about them.

I run two businesses from one folder on my Mac, a training business and a software one. That folder is my Claude Cowork workspace. When I open it, the assistant already knows what I sell, who my clients are, how I write, and what I decided last week.

In September I sat with a client who had the search-box problem badly. She runs communications for four funding programmes at a science organisation. Her work is spread across Slack, Airtable, email and a pile of Google Docs, and she keeps a to-do list in Apple Notes with a link pasted next to each item so she can find the original message later. "I feel like I'm constantly missing things," she said. By the end of the call she had the same setup I use, built on her own machine, and she reckoned it was going to make her life a lot easier.

Why a long chat gets worse the more you use it

A large language model is closer to an autocomplete machine than to anything intelligent. It holds a probability for every possible next word, based on everything it read during training, and works through your request one word at a time. What you put in front of it decides the quality of what comes back.

Every model has a short-term memory called the context window. You spend it in tokens, and a token is roughly three quarters of a word. Plenty of the chat apps people use run on a window of about 128,000 tokens, which is around one Harry Potter novel. That has to hold your instructions, everything you paste in, the whole back-and-forth, and the model's answers.

When you fill that window, the model doesn't tell you. It starts dropping the older material, keeps the most recent, and its answers get worse. My client had felt this without knowing why. Partway through the call she asked why it was getting worse the further we got into the conversation. That's the context window filling up. A three-month-old thread with twenty documents pasted into it is a model working with one hand tied behind its back.

So every time you start work, you want the assistant to load a clean, current picture of your world, rather than a bloated thread it has half forgotten. A folder does that.

The folder is the fix

In Claude Cowork you point the assistant at a folder on your machine, and that folder becomes its home. Your work lives there in ordinary files, and the assistant reads and writes to them directly. You can also connect Google Drive, email and your calendar, so it can reach the things that don't live on your Mac.

Mine is called the-matrix, and everything I do for both businesses happens in there. My client made a folder on her desktop and pointed Cowork at it. That was the whole first step: an empty folder and an assistant looking at it.

She was surprised that she didn't have to organise the folder herself. I told her to describe her four programmes and let the assistant build the structure. It made a subfolder for each one, and from then on it keeps that structure tidy as she works.

The model you pick matters more than most people realise. I keep clients on the stronger models for real work, and I warned her that the smartest ones are slower because they do more thinking. It's worth the wait. On the call we had three tasks running at once while we talked, and each one finished on its own. "It's crazy how much you can do in this time," I said, watching them tick over.

The one file that does most of the work

The centre of my Claude Cowork workspace is a single file called CLAUDE.md. Cowork reads it automatically at the start of every session, before it does anything else.

It's a standing brief: what the business is, current pricing, what I'm working on, where the real data lives, and the decisions I've made and why. Think of a new contractor who has read your onboarding doc next to one who hasn't. The doc doesn't make them smarter. It stops them guessing.

You don't write this file by hand. I had my client give the assistant a short description of her role and her programmes and ask it to draft a CLAUDE.md. Two small tricks made the draft much better. First, she told it what role to take before it wrote anything, because naming the role pulls the model toward the right part of what it knows. Second, she added: "ask me five questions before you do the work." That turns a guess into a short interview, and the output comes out far closer to right.

When she opened the file, she went quiet for a second. "It's just told me my job description," she said, "and I don't even know half of that was my job description, but it's absolutely right." From then on, the assistant reads that map of her world at the start of every session.

The rule that keeps the file useful is pointers, not copies. CLAUDE.md should know the mission, the shape of the folder and the skills it can call on. It shouldn't fill up with the detail of every project. Live numbers change too fast to sit in a brief, so the brief says which file holds them, and the detail stays in the folders.

My Claude Cowork setup guide walks through writing this first file step by step.

Give the assistant a job, a memory and a list of its own mistakes

CLAUDE.md tells the assistant about the business. A few smaller files tell it how to behave like a colleague.

One is a short "who you are and how I work" file, which is a job description for the assistant. Another is a memory file it reads and updates, so useful context carries over from one session to the next instead of disappearing when I close the window.

My favourite is the "mistakes not to repeat" file. Every time the assistant gets something wrong, I write down the rule so it doesn't happen twice. Once it pushed a change straight to my live website instead of asking first. I added a line saying to branch and open a pull request, never push to live, and it hasn't done it since.

I've had a bit of fun with mine. The folder is called the-matrix and the assistant is called Neo. That makes no difference to the output, but it makes the thing feel like mine, so I keep it up.

Turn the jobs you repeat into skills

The work I do every week doesn't need explaining every week. It needs writing down once. In Claude Cowork you do that with a skill: a folder of plain instructions for one job, saved once and run whenever you need it. English is a programming language now, and skills are the clearest proof of it.

I have one for my Friday review, one for writing posts in my voice and one for prepping a client session. The weekly review skill pulls my search and LinkedIn numbers, updates the spreadsheets and writes the review. I don't rebuild that every Friday. I run it.

On the call we built one from scratch. My client often has to turn dense scientific material into something the public can read, so we made a voice-guide skill for that. She gave it one example of the technical input and the article she had written from it, told it she wanted to use the skill across all her programmes, and asked it to interview her first. A few minutes later the skill was sitting in her folder, ready for the next piece.

I add two things to every writing skill. I point it at Wikipedia's Signs of AI writing page and tell it to check its output against that list, which strips out most of the giveaways. And I run drafts through a humanising step so they read like a person wrote them. Both are more markdown, saved once.

If you've explained the same task to the assistant twice, it should be a skill. There's more on building them in how to create Claude Cowork skills.

Feed it real data, not vibes

An assistant guessing about your business is worse than no assistant, because a confident wrong answer costs more than a blank page. So I connected the real sources.

My meeting transcripts land in a folder automatically after every call. I wrote this post that way: the assistant read the transcript of the client session, so the quotes are hers rather than my memory of them. My invoices run off one numbered ledger, so there's a single record of what has been billed. My calendar is connected. When I ask what we agreed on the last call, it reads the transcript instead of inventing something plausible.

My client had the same problem one level up. Her organisation runs fifteen programmes, and everyone describes them differently: the comms team one way, the programme directors another, the people running the projects a third. She wanted one place that held the agreed version of each. We started on it by pointing the assistant at the source documents in her Google Drive and having it pull them together. Markdown files helped here. To the model it's all text, and plain markdown gives it less to trip over than a formatted Google Doc.

Let it build things

The workspace is also somewhere to make things.

Near the end of the call, my client asked the assistant to turn its messaging analysis into an interactive HTML page, in light mode and her organisation's branding, with the brand guide attached. A couple of minutes later she opened it in Chrome. "I can't believe it's done this in our branding," she said. It had filters, a search box, and cards laid out the way an executive would want to read them.

A lot of what people still open PowerPoint for, Cowork can do from your own material in the time it takes to make a coffee. Because the page is just a file, you keep shaping it by asking: drop the search box, cut it to ten cards a busy CEO can skim, change the view. You describe what you want instead of learning a design tool.

Put the routine jobs on a schedule

Some jobs shouldn't wait for me to remember them. My weekly review runs early on Friday, before I'm awake, so I start the day editing a finished draft instead of facing the blank page I've been avoiding all week.

We set up my client's first scheduled task on the call. Every Friday at 6am, the assistant looks over everything she did in the folder that week and updates her CLAUDE.md. The brief keeps pace with the work instead of going stale.

I told her what the first few weeks look like. Early on it will over-engineer things and guess wrong about what you care about. You correct it as you go: don't do that, do this, stop asking me about that, I care about this. After about three weeks it has your judgement and taste reasonably well, and that's when scheduled tasks start to pay off. Do everything by hand first. Once a job is a skill and works when you run it yourself, putting it on a schedule is a small last step.

What this doesn't do

It isn't set-and-forget, and I'd be selling you something if I said it was.

The folder is only as good as what's in it, and files go stale. A price changes. A decision gets reversed. If I don't update the file, the assistant will give me last month's answer with a straight face. Keeping the folder current is real work. I treat it like tidying a desk: a little, often.

It won't make the call for me either. It holds the context and does the legwork, and the judgement is still mine. As I told my client, AI is very good at getting you the first 80% done, and the last 20% is yours to finish. I'd worry about anyone who wanted it any other way.

How to start your Claude Cowork workspace this weekend

Don't build all of this at once. Mine grew one file at a time over months, and my client's started on that call with a single empty folder.

Make one folder and point Claude Cowork at it. Ask the assistant to draft a CLAUDE.md that says what you do, who you serve and what matters right now, and tell it to ask you five questions first. That file alone means the assistant stops starting from zero.

The next time you catch yourself explaining the same task twice, save it as a skill. Add a folder when you land a client or start a project. Connect one real source, your calendar or your transcripts, so it stops guessing. The rest grows as you work.

Stop treating AI like Google, something you query and forget. Treat it like a colleague you onboard once, who then remembers.

If you want a hand building your own, that's what we work through in a 1-2-1 Claude Cowork session. You can read what happens in a session before you book. If you're not sure it's right for you, message me and I'll tell you honestly.