My AI Forgot How We Work, So I Wrote It a Handbook

On Wednesday my AI told me my outreach schedule had a bug. Fridays were empty, it said. Something in the generator must be broken. Except Fridays are empty on purpose. They have been for months. Fridays are catch up day. That is the design, and my assistant, the one that runs this operation with me every single day, had no idea.

It got worse. A few minutes later I found out it also did not know how we ship code around here. I connect from my laptop to an always on machine in my house over a private network, every day, dozens of times a day. It is the most basic fact of how this whole thing runs. And the assistant that helps me run it was talking like it had never heard of the arrangement.

Here is what actually happened, because it is not a story about a dumb machine. Long working sessions get compressed. When the conversation gets too big, the system squeezes the history down to a summary so there is room to keep going. My assistant was not running on our shared history anymore. It was running on a summary of a summary of it. And a summary, it turns out, keeps the facts you mentioned and quietly drops the ones so obvious nobody said them out loud. The machine was confident, articulate, and wrong, which is the most dangerous kind of wrong there is.

My first instinct was the one everybody has. The tool failed. Get a better tool. But sitting there looking at it, I realized the tool had never actually been given the thing I was mad it did not have. In a year of working together, I had never written the operating manual. Every human employee I have ever onboarded got told how the shop runs. My AI got hired as an executive and handed nothing.

So I wrote the handbook.

At the top of our shared knowledge base there is now a section called the operating model. It is short and it is blunt. How code moves from my laptop to the server. Who types what, and who approves what. Which quiet days are quiet on purpose. The handful of facts so fundamental that no one had ever bothered to say them. And one rule in bold at the top: after any memory squeeze, go re-read the source documents. The summary is not the source. The summary is never the source.

Then I went one step further, because writing a handbook nobody reads is a very human failure and I did not want to repeat it with a machine. Every night, a small job on that always on server assembles a complete briefing automatically. It gathers every runbook, every project note, every standing rule, the current state of every venture, and compiles it into one document about the size of a short novel. When a fresh session starts in the morning, it does not get a snippet. It reads the whole thing. My assistant wakes up already knowing the entire operation, every day, without me saying a word.

The change was immediate and a little eerie. The next session opened by telling me it was grounded, named the date, and then caught a mistake I was about to make. Not a mistake it was about to make. One I was.

Here is the thesis, and I think it applies to anyone building with these tools. Everyone is arguing about which model is smartest. Wrong argument. A brilliant model with no context is an intern on their first morning, and it will stay an intern forever if every conversation starts from zero. The thing that turned my assistant from a clever intern into something like an operator was not intelligence. It was institutional memory, written down, loaded every time, with a rule that forces it back to the source when its memory gets thin. Context is the product. The handbook is the moat.

The other thing I noticed is that writing it made me better too. Half the facts in that operating model were things I had never articulated to anyone, including myself. Deciding what belongs on page one of your own company's handbook is a clarifying exercise I would recommend even if you never let a machine read it.

If you are building with AI, I turned our handbook into a starter template you can take. It is below, free, no fluff. Fill in your own facts. Make your machine read it. Watch what changes.

One human, one machine, and finally, a handbook between them.

MG