Ryan “RyMac” McKinney · USMC · 16 August 2026 · 8 minute read
Most folder structure advice assumes you run one thing. Agencies never run one thing. You have your own brand, 6 client brands, and a side offer you keep meaning to launch, and every one of them has different rules. Here is how to split them so they stop bleeding into each other, and the 5 minute audit that catches it when they start.
The structure below, doing a real job start to finish. One night, one agent, one folder system, and a page that converted at 28% on the first 89 readers. No cuts where it went wrong.
Give every brand its own folder holding only what is different about that brand. Keep the rules that are the same for all of them in one shared folder, and have each brand point at that folder instead of copying it.
Anything you copy into 2 brands will drift apart. That is not a discipline problem. It is arithmetic.
I read the whole first page of Google for this. Wikipedia, a Harvard data management page, a Box admin doc, 3 posts about organising a codebase, and a Reddit thread where people share screenshots of their Documents folder.
Every one of them is written for a human who needs to find a file.
Not one is written for the situation you are actually in, which is that an AI is reading these folders and making decisions from them. A human who opens the wrong folder notices and backs out. An AI opens the wrong folder and confidently does the work.
That difference is the whole reason an agency folder structure needs designing rather than growing.
Before you move a single file, answer one question about every rule you have.
Is this true for every brand, or only this one?
That is it. That question sorts your entire operation into 2 piles, and the 2 piles get treated completely differently.
Same for everyone goes in one shared file. Your invoicing process. Your tone rules. How a job gets approved before it ships. One copy, and every brand points at it.
Different per brand goes inside that brand's folder. Their colours. Their voice. Where their traffic goes. Who signs off.
Get this backwards and you get the 2 failures every agency folder eventually has: the same instruction written in 6 places slowly disagreeing with itself, or 6 brands sharing one instruction that was only ever right for one of them.
This is the shape, and it is smaller than people expect.
~/agency/ ├── CLAUDE.md <- the map. points at brands. never does work ├── Context.md <- the desk. what we're doing this week ├── shared/ <- TRUE FOR EVERY BRAND. one copy only │ ├── how-we-work.md │ ├── voice-and-bans.md │ └── approval.md └── brands/ ├── acme-roofing/ │ ├── CLAUDE.md <- what's DIFFERENT about Acme │ ├── brand.md <- colours, voice, who signs off │ └── work/ ├── northside-dental/ └── our-own-brand/
3 things are load bearing here and none of them are the folders.
The map at the top does no work. Its only job is to say which brand folders exist and that shared/ holds the common rules. Start putting instructions in it and it becomes one more thing that gets read on every single request. Same principle as the 11 line map for a single business.
Each brand folder holds differences only. If acme-roofing/brand.md repeats your invoicing process, delete it and point at shared/. The test is brutal and simple: could this line be identical in another brand folder? Then it does not live here.
Nothing is copied. Everything is pointed at. A pointer cannot go stale in one place and stay fresh in another. A copy always does.
This is where the $5,000 course crowd will tell you to spin up a specialist agent for each brand and have them coordinate.
You do not need that, and the research is not on their side either. Microsoft's own guidance on choosing between one agent and several says the same thing in enterprise language: more agents means more coordination cost, and you take that on only when the work genuinely cannot be done in one place. A 2025 paper comparing the 2, Single-agent or Multi-agent Systems? Why Not Both?, puts numbers on the same trade: several agents cost more to run and more to keep working.
6 brands is not 6 kinds of work. It is 1 kind of work, 6 times, with different details.
One worker who can read which folder they are standing in handles that. 6 workers who have to tell each other what they did is a coordination job you invented and now have to manage. That is the difference between a filing problem and an agent memory problem, and almost every agency I talk to has the first one.
Structures do not announce when they break. Run this monthly, or any time the work starts feeling harder than it did.
1. The standing test. Open a terminal inside one brand folder and ask a question only that brand can answer. What colours does this one use, where does its traffic go. If the answer comes back right, the structure holds.
2. The question test, and this is the sharp one. If your AI asks you something the folder already answers, the structure is broken, not the AI. Being asked is the alarm. It means whatever is reading the folder cannot tell where it is standing, and it is about to get everything downstream wrong without telling you.
3. The 2 copies test. Pick any fact that matters. A price, a colour, a status. Search every folder for it. If it appears in 2 files you have already lost, because one of them is stale and nobody knows which.
4. The wrong brand test. Search each brand folder for another brand's details. An agency folder that has been running a while always has at least one leak.
5. The stale desk test. Open Context.md. If it describes work you finished last month, everything reading it is being handed a map to nowhere.
I ran this on my own setup the day I wrote this, on camera, and it was not clean.
I run 2 public brands out of one workspace. Their rules had been sharing a single file for months, and that file had quietly grown to serve 5 different things at once. Neither brand could get a straight answer out of it.
The 2 copies test caught the real one. The same fact was written in 2 files and they disagreed. One said a video was live on the site. The other said it had never been added. Both were written by me, weeks apart, and I believed whichever one I opened last.
The fix took an afternoon and it moved no files at all. The rules that were genuinely different went into each brand's own folder. The 40 or so rules that were the same for both stayed in one file that both point at. Nothing that was already working got touched.
That is the part worth stealing. Splitting brands is a rules job, not a moving job. If your plan involves dragging folders around for a day, you are solving the wrong half.
A brand folder that grew a copy of the shared rules. Usually because someone needed a small change and copied the whole file to make it. Point at the shared file and write only the exception.
Client work filed by date instead of by brand. Now nothing about that client is in one place, and every question needs 4 folders open.
A brand folder for a client you signed and never started. An empty folder reads as a real one. Do not create it until there is work in it.
Naming folders after people. People leave. Brands do not.
Letting the desk file rot. This is the one that gets everybody, and it is the same failure whether you run 1 brand or 20. Update it before you stop working, not when you next need it.
The method is Interpretable Context Methodology, ICM, and it is not mine. Jake Van Clief and David McDermott wrote the paper: folder structure as agent architecture. Jake teaches it himself over at Clief Notes. What I do is run it in public on real brands and write down where it bends.
If you want the method explained properly rather than applied: Interpretable Context Methodology in plain English. The idea underneath all of it is context engineering, and the single brand version of this structure is the ICM folder system.
I owned a 7 figure agency and trained agencies for years before any of this existed. The filing problem is the same problem it always was. The only thing that changed is that something else reads the files now, and it does not ask for directions.
CLAUDE.md, Context.md, the folder skeleton, a 1-page setup sheet, and a teardown checklist for whatever you are already paying for. 20 minutes, start to finish. No card, no call, no course to buy.
Send me the folder system