The AI Architecture Layer (aka AI Brutalism)

Layer 1: Architecture. The plumbing. Where the AI sits, how it's wired: folder structure, project brains, and the maintenance loop that keeps it all current.

The strongest AI setup is brutalist. Ugly, structural, no shiny bits, just folders and markdown doing the work that looks like magic from the outside.

Brutalism is the architecture of raw concrete and exposed structure. No cladding, no decoration, nothing hidden. The building is honest about how it's built, and people either love it or call it ugly. The best way to wire AI into a business looks exactly the same.

Last week I said skills are the last layer. A few people asked the obvious next question. If skills are last, what's first? This is. Architecture. Where the AI actually sits inside your business, and how it's wired in. The least glamorous layer, and the one that decides whether everything above it works. And in this context, architecture quite literally means your folder structure and your filing system. That's it.

The problem it solves shows up the moment you start. Most people open Claude or ChatGPT from zero every single time. They re-explain who they are, what they sell, how they talk, who the customer is. The model is sharp for that one conversation and then it forgets all of it. Tomorrow you do it again. You're the bottleneck, retyping your own business into a blank box forever.

Architecture kills that. You build the business once, as files, in a structure the model reads before you type a word.

Mine is just folders. One for each part of the business. Works. The community. The clients. Operations. Content. Nothing clever, just the actual shape of what I run, written down where a model can walk it.

At the root sits a single file. A CLAUDE.md. It's the map. It tells the model what lives where, the rules that never bend, and where to go for any given job. Then inside every project folder sits another one. A CLAUDE.md scoped to that project, holding the context that only matters there. The client's situation. The state of play. How they like being worked with.

So when I open a conversation about a specific client, the model isn't guessing. It reads the root map, walks to that client's folder, loads that client's context, and shows up already knowing the account. I've built a specific rule so I can talk to it in natural language. I just say "fire up works" and the LLM loads all the context.

Two things come out of that. The first is persistent context. You build it once and it's there every session, so the model never starts from scratch again. The second is scope. Each significant project gets its own folder, so the AI loads only what that job needs, not every file across your whole operation. That's the actual skill. Managing context. Not too big, not too small. The Goldilocks of context. Hand it your entire business to write one client email and it drowns in noise. Hand it nothing and it guesses. The folder structure is how you give it exactly the right amount, every time.

That's the whole idea. The model starts every conversation as a senior operator who already knows the business, instead of a stranger you have to brief.

People build this once, feel like a genius for a week, and then never touch it again. Three months later it's a museum. The pricing changed, the positioning moved, the client churned, and the files still say otherwise. Stale context that looks authoritative is worse than no context. It's how a confident model gets it confidently wrong.

So the architecture has to maintain itself, and that's a job for two skills, not for discipline. At the end of a working session I run /debrief. It writes the decisions and changes straight back into the project files. Then /reflect captures what I learned about how to work better with the files, and saves that too. The business changes, the files change with it, including the skills I run. It happens automatically, a side effect of doing the work. The architecture stays alive because updating it is part of the loop, not a chore I have to remember. That's compound learning. It's where you really start to power ahead with AI.

This approach works as a personal setup with purely local files, but the same structure scales to a whole company. A shared brain the entire team builds on, instead of one person's folders. That's a whole other post, but the short version is you put the folders somewhere the team shares. GitHub, Dropbox, Google Drive, OneDrive, they all work. The one requirement is that you can mirror the files locally.

This is why at Works, this is where we start every client. Not with a shiny agent. The agent everyone wants, the one that runs the business while you sleep, is sitting at the top of a stack. Underneath it is the skill. Underneath the skill is the context. Underneath the context is this. The folders, the CLAUDE.md files, the maintenance loop. Build it first and everything above it gets sharper for free. Skip it and every clever thing you build on top is standing on sand.

It doesn't demo well. It's folders and markdown, raw and exposed, the brutalist layer of the whole stack. No animation, no viral screenshot. It's what decides whether your AI grows into a colleague that compounds, or never gets past a party trick.

Three things to do this week if you're running AI in your business.

First, make one folder for your business and drop a CLAUDE.md in it. Five lines. What you do, who you serve, how you talk, the rules that never bend.

Second, give each real project its own folder and its own CLAUDE.md. Scoped context beats one giant file every time.

Third, decide how it stays current before you need it to. A skill, a habit, a Friday ten minutes. Pick something, because the version of this that goes stale is the version that quietly makes your AI worse.

The folders are the boring part. They're also the part that does the work.

This is the first layer we build at Works. Architecture, then context, then capability, then skills on top. If your team has a pile of AI tools and nothing sticking, it's almost never the tools. It's that there's nothing underneath them.

Let's get to Work!