Could your team share an AI assistant?

Most people in a business now have their own AI. Some use the free chat window, some pay for one, and a growing number have moved on to the tools that do work rather than answer questions: Claude Cowork, ChatGPT Work, Claude Code, Codex. What almost none of them share is anything the assistant has learned. Each one starts the morning knowing nothing about the company, learns a little from whoever it is talking to, and forgets it when that person goes home. Two things fix that, and neither is a bigger model.

Why does my AI keep forgetting?

Because memory, in the chat tools, is a nice-to-have bolted onto a conversation. ChatGPT, Claude and Gemini will all remember some things about you now, and it does make them more useful. It is also thin. It is the assistant’s own notes about you, kept in its own way, and you cannot really see them, edit them or hand them to anyone else.

The tools that do work take a different approach. Claude Code and Codex both read a plain text file before they start anything, and the office-side tools built on them, Claude Cowork and ChatGPT Work, are heading the same way. Ordinary documents, in a folder, that you can open and change. Think of the folder a good new starter builds in their first month: what the company does, who is who, how we do things here, what the boss means when they say “just tidy that up”. Written down once, read every morning. That is what these files are, except the new starter is the assistant and it reads the whole folder before every task.

For one person, that folder can get quite rich. What the business is. How you like to work. How you like to be spoken to. What you like, what you do not, and what to do when you say certain things. The assistant becomes yours in a way a chat window never does, because everything it knows about you is in a form you wrote or approved.

What is a harness, in plain English?

That folder of files, plus the rules for keeping it up to date. People who build these things call it a harness, which is not a helpful word, so here is the plain version: it is the set of instructions the assistant reads before it does anything, and the routine that improves them.

Claude Code, for one, lets you attach small checks that run at set moments: before a file is saved, after a task finishes, when you correct the assistant. So when you say “no, we never send an invoice on a Friday”, that correction does not vanish into the conversation. A check catches it, writes it into a lessons file, and next morning the assistant reads the lesson along with everything else. Correct it once and it stays corrected. Over a few months you have an assistant that has been trained by you, on the job, in a way you can read back.

We run ours this way. The assistant that helps write these articles reads a set of plain text files every session: what the business does, how it writes, what it may and may not do without asking, and the lessons it has been given. Nothing about how it behaves lives anywhere I cannot open.

That is one person and one assistant. Everything from here is what the same files make possible across a team, and I have not run it across one yet. Read it as a design worth piloting, not a report on one.

What happens if we share the files?

The AI Daily Brief spent an episode this month on what it called the shift from single-player to multiplayer AI: Anthropic’s own teams running assistants that belong to a Slack channel rather than to a person, and coding tools that let two people work in one assistant’s session at the same time. The shared assistant is the headline. The shared files underneath it are the useful bit for a business of ten or fifty people, because you already have them.

Put some of those files in a folder everyone’s assistant can read, and two things become possible straight away.

The first is shared lessons. Every assistant in the business learns corrections from its own person. Overnight, an editor job reads the lessons that have been approved for sharing, spots the ones that keep coming up, and writes them into the shared set once. The lesson one person taught their assistant on Monday is in everyone’s assistant by Tuesday, without a memo.

The second is shared expertise. Each person’s assistant, as it works alongside them, can keep a plain text note about that person’s area of the job: how the month-end runs, what the client onboarding checklist actually involves, which supplier needs chasing and how. When somebody else in the business has a question about that area, their assistant can read the note and answer it. No meeting, no email, no waiting for the one person who knows to be back from lunch.

Once the files are in one place there are dozens more of these, and the businesses that find them will be the ones that started with two.

Which files should stay private?

That is a business decision, not a technical one. Tone of voice stays yours. A lesson about how the company raises an invoice belongs to the company. Between those is a wide band that depends on the culture of the place, and some of what one person’s assistant knows is specific to one client or one piece of work and should not travel across the business at all.

One pattern that seems to work is two lessons files per person. The general one stays private, and the assistant keeps it. Alongside it, a shorter file where the assistant writes down the lessons and notes it thinks are about the business rather than about the person. Nothing in that second file moves on its own. Each morning the person is shown the list of what their assistant thinks should be shared, and says yes or no to each line. The assistant suggests, the human decides, and the line is visible every day rather than argued about once.

Write the line down too. Not as a policy document, as a short list: these files are shared, these are not, this is who can change the shared set. The businesses that skip this step will not have a data problem first. They will have a trust problem, when someone finds a note about their work in a folder they did not know existed.

Won’t it just read everything?

It will if you let it, and this is the mistake to avoid. An assistant that reads every file before every question does not get wiser. It gets slower, then vaguer, then wrong, because the one fact it needed is buried under fifty it did not.

The fix is a contents page. A short index at the top of the folder, one line per file, saying what it is for. Assistants read the index first and open only the file that answers the question. Inside each file the discipline is the same: name the thing, state the fact, point at where the detail lives. A note that says “Month-end: closes on the third working day, checklist in the finance folder, ask Priya if the bank feed is late” does more work than a page of prose about month-end, because the assistant can read the whole thing in a breath and knows where to go next.

This is also why the overnight editor matters more than it sounds. Its job is not only to spread lessons. It is to keep the shared set short: merge the duplicates, retire what has stopped being true, and refuse to let the folder grow into the thing nobody reads.

Where does the shared folder live?

On storage the business controls. This is the part where data protection stops being a footnote, because the notes in that folder will mention people, clients and money, and every assistant that reads them sends the contents to a model somewhere.

Three things settle most of it. First, keep the folder on your own systems or a UK-hosted service you have a contract with, not in a free tool’s cloud. Second, only let it be read by assistants on a paid business tier with a no-training clause and a data processing agreement, or by a model you run yourself. Third, strip what the model does not need before it goes in: a note about how a client is onboarded does not need the client’s name, and a lesson about invoicing does not need the invoice number. Those three rules leave the files no more exposed than the documents they were written from.

There is a quieter reason to own the folder. The knowledge in it is the business’s, and for the first time it is written down somewhere other than the heads of the people who hold it. When someone leaves, their assistant’s notes about their job stay. That is worth more than most of the automation people are buying, and it costs a folder.