Stop Trapping Your AI in One Chat App
Different agents, different apps, different models, all walking into the same room and knowing where they are. The folder is the room.
The idea
A folder harness is an ordinary folder of plain text files that tells any AI agent where it is, what you are working on, and how you like to work, before it does anything else.
Different agents, running in different apps, on different models, can all enter that same folder and pick up where the last one left off. They do not share a chat history. What they share is the folder.
That is the whole trick, and it is why my AI system does not live inside one company’s chat app.
The model changes. The room stays the same.
The rest of this is the five parts of that room and what each one is for.
The front door
The boot file tells the agent what room it woke up in
AGENTS.md is the first thing a new agent reads. It is the boot file.
It does three jobs. It tells the agent where it is. It tells the agent how to work with me. And it points at deeper context instead of trying to hold everything itself.
If you were around for DOS, this is config.sys and autoexec.bat. Those files did not do the work. They got the machine into a known state so the work could start. The boot file does the same thing for an agent: it gets it into the room and oriented, and then gets out of the way.
The mistake is letting this file swell. Everything you cram in here gets read on every single session, whether it matters today or not. A boot file is a router, not a warehouse.
Right now
The checkpoint keeps the next session from waking up cold
session-state.md holds what is happening right now.
This is the sticky note I leave on my desk at the end of the day, telling tomorrow’s version of me what needs handling. It says what is active, what just happened, and what the next useful move is.
Without it, every new chat starts from scratch. With it, the next agent walks in already knowing the thread.
Current state is what is happening now. Memory is what should still matter later.
That line is worth holding onto, because it is the difference between the two files people most often collapse into one.
The rooms
Projects keep separate work separate
If the checkpoint keeps the current thread alive, the projects/ folder keeps each body of work from bleeding into the others.
Every real project gets its own room. Launching a YouTube channel. A family TV picker. Website builds. A Mustang restomod out in the garage. Those are all real things I am working on, and none of them should have to know about each other.
Two things fall out of that. An agent can re-enter any one project without me re-explaining the whole history. And a big project can grow without polluting the boot file that every session has to read.
What sticks
Memory keeps the lessons from disappearing
Projects keep the work organized. Memory keeps what you learned from evaporating when the chat closes.
Memory is where durable context lives, split by what it is for:
- Decisions, so settled things stay settled
- Corrections, so the same mistake stops repeating
- Preferences, so the agent stays calibrated to how you work
- References, so useful external things are findable again
The agent needs one durable place to look. Not a search through six months of conversations, which is exactly the thing that does not work.
The routines
Skills stop you re-explaining the same process
Some work is repeatable. Publishing a video. Running a review. Checking a build before it ships.
A skill is that workflow written down once as a file, so the agent can read it when the job comes up instead of getting the same lecture from you every time.
Here is the part I find genuinely interesting: this is how the system gets better without the model getting better. You are not waiting on a smarter release. You are writing down what you already figured out.
A lot of people are talking about skills in one specific tool right now. The idea is much bigger than any one tool. Save the workflow once. Let whatever agent shows up read it.
The payoff
The agent can change, because the room does not
Put those five together and you get the thing that made me stop worrying about which AI I use.
Codex can read the folder. Claude Code can read it. Cursor can read it. Whatever launches next month can read it. If an agent can open a folder, it can use the context stored there.
This only works with an AI tool that can read files on your computer, or one you can hand the files to directly. An isolated browser chat cannot walk into a folder on its own.
That is the whole cost of entry. There is no install, no database, and no subscription attached to the folder itself.
The model does not compound by itself. Each new version arrives clever and empty, knowing nothing about your work.
The written environment around the model is the part that compounds.
The chat ends. The app changes. The model gets replaced. The folder keeps the rules, the current work, the decisions, the projects, and the workflows.
That is the harness.
Your turn
Start with the front door
You do not need all five parts to feel the difference. Start with the one that does the most work.
Make a folder you control, and put one file in it:
# Who I Am
A few sentences about you and the work you actually do.
# What I Am Working On
The projects that matter right now, one line each.
# How To Work With Me
How you like information presented. What you keep asking for.
What you keep asking the AI to stop doing.
# Where Things Live
Point at the other files as you add them. This file routes.
It does not hold everything.Then open a file-capable agent in that folder and tell it to read AGENTS.md before doing anything else.
Add the next file when you feel the need for it, not before. The checkpoint when you get tired of re-explaining yesterday. The memory files when the same correction comes up twice. Skills when you notice you have described the same process three times.
Built that way, the folder stays yours, and so does everything in it.
Watch it
Stop Trapping Your AI in One Chat App
The video walks the same five parts on screen, in my own voice, with the real folder open. This page is the reference version so you can look up one piece without scrubbing the timeline.
