I Stopped Chatting With AI and Gave It a Folder Instead
A chat ends. A written handbook survives. Here is how I started giving each new AI session the context it needed without explaining everything again.
The idea
Giving an AI a folder means writing down the context that should survive the current chat, then letting a file-capable AI agent read those files before it starts work.
That is the whole idea.
The folder is not special software. It is not a database or some hidden AI memory system. Mine is an ordinary folder on my computer filled with plain text files.
Those files tell the next AI session who I am, what I am working on, how I like to work, what we have already decided, and which mistakes we do not need to repeat.
Six months ago, I started working this way because regular AI chats kept forgetting work I knew we had done together. That one change did more for me than any prompt trick I had learned.
The problem
Chat history is not a handbook
Earlier this year, I was preparing for a quarterly meeting at work.
I had used ChatGPT almost every day for months. It had helped me think through projects, solve problems, and create things. I assumed I could ask it a simple question:
What have we worked on over the last four months?
It could not tell me.
The work existed, but it was scattered across separate conversations. Each chat knew what happened inside its own window. None of them held a reliable record of the larger body of work.
I had treated conversation history like company documentation. It is not.
A new employee does not learn the job by combing through months of old Slack conversations. They need the current handbook, the active project notes, the decisions that have already been made, and the procedures the last person figured out.
So I wrote a handbook for my AI.
What is inside
What goes in the folder
My folder grew far beyond one file, but the material inside it answers a small set of practical questions:
- Who am I?
- What am I working on?
- How do I prefer to work?
- What have we already decided?
- What has gone wrong before?
- Which processes should happen the same way next time?
The answers are stored as plain text. Some files describe me. Some track current projects. Others preserve decisions, corrections, or reusable procedures.
The AI model does not absorb these files permanently. The next session reads them when it enters the folder. That distinction matters.
The model is still replaceable. The handbook stays with the work.
If I switch from one file-capable AI agent to another, I do not need to recreate my entire working history inside a new chat. I give the new agent access to the same folder and tell it where to begin.
The proof
The app that proved the problem
The problem became impossible to ignore while I was building my first real application with AI.
The app handled orders at work. The old process involved FTP transfers and tiny CSV files. It worked, but it looked and behaved like something from another generation of computers.
I described what we needed, and Codex helped me build a visual, usable application around that process. For somebody who does not write code, seeing a real application come together from plain-language instructions was a hell of a moment.
Then I would open a new session the following day.
The AI that had deployed the application yesterday suddenly acted as if deployment were my responsibility. It no longer knew the process it had already used. I would point out that we had done this for several days in a row, but that did not put yesterday’s working context into today’s session.
It was the smartest new hire in the building, arriving for work every morning with no handbook.
The fix was not a more elaborate prompt. Once we solved something, I had the AI write the procedure down. The next session could read the same instructions instead of rediscovering the process or handing the job back to me.
The application did not become maintainable because the model magically learned from me. It became maintainable because the working knowledge stopped living only inside temporary conversations.
Your turn
Start with one file about you
You do not need my full folder to test the idea.
Create a new folder somewhere you control. Open Notepad or another plain text editor and create one file called USER.md.
The .md extension means Markdown, but you do not need to learn Markdown first. It is still a plain text file. Notepad counts.
Put this inside:
# About Me
## Who I Am
Write a few sentences about who you are and the kind of work you do.
## What I Am Working On
List the projects or goals that matter right now.
## How I Prefer to Work
Explain how you like information presented, how much detail you want,
and anything you repeatedly ask an AI to do or avoid.Do not turn this into an autobiography. Give the next session enough context to stop treating you like a stranger.
Then open a file-capable AI agent inside that folder, or give the agent access to it, and say:
Read USER.md before we begin. Tell me what you understand about me,
what I am working on, and how you should work with me.An isolated browser chat cannot automatically inspect a folder on your computer. This requires an AI tool that can read local files, or you must attach the file yourself.
Keep private information out of the file unless you are comfortable sharing it with the AI service you use. Plain text makes the system easy to inspect, but you still decide what belongs in the handbook.
That first file will not run an entire AI system. It does something smaller and more useful: it proves that durable context can live outside the chat.
The chat can end. The next new hire can still read the handbook.
Watch it
I Stopped Chatting With AI and Gave It a Folder Instead
The video includes the full story in my own voice. This page is arranged as a reference so you can return to the explanation and build the first file without scrubbing through the timeline.
Read next
