The ICM Method That Upgraded My AI Operating System
A numbered filesystem gives your agent memory you can open and check yourself. This is how it works.
My business runs inside a folder called AI Operator OS. It is my second brain. Every article draft, client brief, corrections log and skill my agent uses sits in there as plain markdown, and it holds years of accumulated context about how this business works.
A second brain is good at storing that context. What it does not do is decide which of it your agent should read for a given job. And it gives you no easy way to see the state of a workflow. You can describe your own system perfectly and still have no idea what the agent actually loaded, or where any piece of work stands, without opening folders one at a time.
That was my position. The fix turned out to be structural, and it came from a method called ICM.
Here's what you're going to walk away with after reading this article:
No Contracts Anywhere
A scan of the vault found almost no CONTEXT.md files anywhere. Six, across the entire system.
Nothing in any folder told the agent what that folder was for. Nothing said what it should take in, what process ran inside it, or what it should hand back. With no explicit contract, the job had to be inferred from the files sitting nearby.
The skills had absorbed the machinery that should have lived next to them. Templates, scoring tables, process rules and examples were packed inside the skill files themselves, so the agent dragged the whole workshop into every session that needed one small part of it.
The root CLAUDE.md had swollen to about 150 lines trying to route the whole vault from one file. Every new rule and every new correction went into it, and the agent carried all of it into every job whether the job needed it or not.
ICM
In 2026, I came across ICM: Interpretable Context Methodology, associated with Jake Van Clief and David McDermott.
It takes a different position from the multi-agent frameworks everywhere right now. ICM uses filesystem structure instead of a framework, and it is built for sequential workflows where a human reviews each output before the next step runs. The agent works one stage at a time, and the folder structure carries the context between stages.
That fit how I already work. I review every output before anything moves on. What I was missing was a structure that made the reviewing easy.
In the last few weeks I installed the icm-architect Claude skill and pointed it at my vault. The rest of this article is what it found, what it built, and how to run the same audit on your own setup.
Numbers Do The Routing
The first idea is almost embarrassingly simple. Folders get numbers, and the numbers do the routing.
A workflow becomes a sequence. A file sitting in an early-numbered folder is early-stage work. A file sitting in a later-numbered folder is further along. The agent does not need to be told what stage it is in, because the folder names tell it.
You can read the state of a pipeline from a directory listing. No dashboard, no database query, no asking the agent where things are up to.
The Root File Is a Router
The second idea is a rule about the root CLAUDE.md. It stays small.
Under ICM the root file is a router, not an encyclopedia. It names the numbered folders, says what each one is for, and stops. The detail moves down into the folders where the work actually happens.
The agent loads the map at the start of a session, then opens only the pages the current job asks for. My 150-line root file did the opposite. It made every job carry the whole vault.
Every Folder Gets a Contract
The third idea is the one my vault was missing entirely. Each working folder gets a CONTEXT.md file that acts as a contract.
The contract says what inputs the folder takes, what process runs inside it, what output it produces, and where the human check happens. When the agent enters a folder, the first thing it reads is the contract, so it never has to infer the job from the surrounding files.
This is what makes review cheap. You are not reading the agent's mind, you are reading its paperwork. When an output misses, the contract is the thing you fix.
Factory and Product Split
ICM separates the machinery from the work.
The factory is the skills, the templates and the process rules. The product is the drafts, the records and the deliverables. With the split, the agent loads only the context for the step it is on. The drafting step loads the drafting context. The review step loads the review context. Everything else stays on the shelf, which keeps each run smaller and far easier to check.
New projects come from template folders. Instead of building a project structure from memory, the agent copies a template, and the copy carries a stamp saying where it came from.
The Skill That Ate My Vault
The main demo was my content-create skill, the one I use for every post and article.
The original file was 768 lines, 8,572 words, roughly 11k to 12k tokens. It auto-loaded five more files on top of itself. A simple hooks request pulled more than 30k tokens before a single hook was written.
The humanizer existed twice. About 14 unnumbered post bullets in one branch, 31 numbered article rules in the other. Rules including the arrow ban, the eyebrow ban, zero rhetorical questions, the "the gap" ban and the one-use-per-phrase rule did not apply to short-form posts.
It contradicted VOICE.md on em dashes and on pass order. It had four stale VOICE cross-references, a stale sponsored-post output path, a dead Voice_Evolution_Report glob, 53 unicode arrows, 96 em dashes, an out-of-order changelog, and unmeasured quality guarantees.
One Router, Two Pipelines
The v2 restructure split it into a router plus post and article pipelines plus shared references.
The current SKILL.md is 55 lines. The shared humanizer is one file loaded by both branches.
A hooks request enters the post contract. An article request enters the article contract. Both use the same humanizer reference.
I am not claiming the rewrite made my posts better, or that it saved tokens I have not measured. The improvement is that I can open the folder and see which contract the agent entered, which reference it loaded, and which rules it applied. That is what interpretable means.
Try It On One Folder
Do not restructure your whole vault. That instinct is the same one that bloated my root file in the first place.
The method and the skill are public. The two repos worth reading are icm-architect and Interpretable-Context-Methodology. Install the skill, then pick one folder where real work repeats. Not a test folder. A workflow you actually run.
Then hand your agent this, with your own path filled in:
Read the ICM skill, then run an inventory-only audit on [your folder path].
Do not move, rename, create or delete any files. Count the records or
projects already there, classify each as complete, partial or empty,
and report what you find, along with what structure this folder would
need under ICM.
Stop after the report. I will decide what happens next.
Inventory first. You decide what happens after you have seen the state of your own work.
The Sponsored Posts Check
The same audit ran on my Sponsored Posts folder as a secondary proof point.
The generated index found 24 records: 9 complete, 7 partial, 8 empty. No existing files moved during the pilot. New CONTEXT.md contracts, template stamps and the generated index were added.
I had been carrying that state in my head, and my head had it wrong. One generated file answered which deals were waiting on me, which were half-finished, and which folders had sat empty since the day I made them.
Memory You Can Open
My second brain gives me somewhere to keep the context. ICM adds a filesystem that tells the agent which context belongs to the job, and lets me open the same files to check it.
I can open that filesystem in Finder, read it in Obsidian, and check the files against the contracts with my own eyes.
When the agent produces something odd, I do not debug a framework. I walk the numbered folders and find the stage where the context went wrong.
If you want help finding the one process in your business worth automating, book a free 30-minute AI Audit: https://theaiarchitects.com/x/audit/icm-article