I stopped repeating myself to Claude. I built an Agency instead.

Every new chat used to start the same way. I'd open the app, sit down, type "right — pull up the vault, read environment.md, remember Sebastian and Mary, here's the project I'm on, the client, the budget, the brand voice…" Twenty minutes of re-briefing before the first useful sentence. So this morning I built a desktop app. One click, and Claude walks in already knowing the room.

The problem nobody warns you about

Twelve days into running AI rather than using it, I have something most people working with Claude don't yet have: a vault. Articulate sits on roughly 30,000 words of structured context — an environment doc with the hard rules I keep breaking, a toolbox with every tool's cost and status, named personas stood up last weekend who each own a slice of the operation, a brand book, a decisions log, a fail ledger. None of it is theoretical. A 267-page research synthesis shipped this weekend that drew on it.

The problem is that none of it shows up when I open a new chat. Cowork — the desktop app I use most — boots blank every session. No folder mounted. No environment.md. No personas. I get the same model, the same skills shelf, and the same blank face that doesn't know I'm British, that ~/ClaudeWork/ is where everything lives, or that "Mary" means a project manager not a holiday.

So I'd type the briefing. Or I'd forget to, hand off a task, and watch the model improvise context it didn't have. The fail ledger has four entries this week alone where the root cause was "started without loading environment.md." Same fail, four times, in a system I'm meant to be running competently.

The fix is a doorway, not a feature

The proper solution is a Cowork setting that auto-mounts a default folder and reads its CLAUDE.md on every new session. That's an open feature request on the Claude Code tracker. Not shipped yet.

The workaround takes ten minutes. A native macOS .app — bundled the old way, by hand — that opens Terminal and runs the Claude Code CLI in ~/ClaudeWork. The CLI version does walk up the directory looking for CLAUDE.md, finds it, loads the canon, picks up the MCP servers I've registered, exposes the named skills, and presents a prompt that already knows me. Same model. Different doorway.

The .app is three files in a structured folder: a launcher, a manifest, and an icon. The icon is the Articulate AI mark — the comet from the wordmark — set on the brand red so it reads at Dock size and at Desktop size. Same shape every other macOS app uses for a doorway. It looks like this:

Double-click. Terminal opens. Claude Code starts in the vault. The first prompt back is Anthony-specific. No briefing. No context tax. Just a senior colleague who's already read the file.

The wow moment — I didn't build it. I watched it get built.

Worth saying clearly because it's the bit that genuinely surprised me. I didn't write a single line of the launcher. I described what I wanted. The Cowork session asked for screen-control permission, I clicked yes, and then I watched the cursor move on its own. Finder opened. The brand folder selected itself. The icon assembled. The bundle wrote itself onto the Desktop. The .app double-clicked itself. Claude Code launched. I had a coffee while it ran.

Then it screenshotted the result, decided it had worked, and wrote me a short message saying it was done.

I've been in this industry twenty-five years. I've seen plenty of demos. I have not previously watched an assistant pick up the keyboard and the mouse and execute a multi-step desktop build, on my actual machine, with my actual brand assets, while I sat back. The recording is going on LinkedIn this week.

The technical bit — Cowork's computer-use permissions, the way Terminal sits in a restricted tier and Finder doesn't — is its own post. The point here is that the same colleague who'd just read environment.md was also able to act on it. Memory plus agency. Not one or the other.

What it feels like

Like walking into my agency on Monday morning.

SuperSebastian is at his desk reviewing the stack. MotherMary has the standup already pinned. StarTrekScotty's in the engine room. The MCPs are wired. The brand book is loaded. The client folders are within reach. I don't have to introduce anybody. The system already knows that Candance has a Monday deadline, that the Christies session is Friday 9am, that the villa corner room decision is owed by 28 May, that I'm two days a week into Articulate Phase 1 and 1 of 11 deliverables ticked. I sit down. They turn round and ask what I want to do.

That's the unlock. Not faster output — the model isn't faster. The unlock is the absence of re-staging. I no longer treat the AI as a forgetful colleague I retrain each meeting. I treat it as a room I walked out of forty minutes ago that's still set up exactly the way I left it.

Why this is the actual product

Every consultancy I've worked in has had a version of this room. Files on the shelf. Senior partners who remember the last campaign. A standing reference to the brand voice. A discipline of writing things down so the next person doesn't relearn. The AI version isn't a different problem. It's the same problem with the file cabinet moved into a vault and the partners moved into named personas. The desktop icon is the door to the office.

The work I do for clients — the six workflows installed on their stack in eight weeks — has the same shape. Pre-staged context. Named agents who own a slice. A doorway the team walks through on Monday and finds the same office on Tuesday. The handover at week eight isn't "here's a tool, good luck." It's "here's the room, walk in tomorrow morning."

The bigger point

The first generation of AI workspaces look like chat windows. Blank slates that learn you over the conversation and forget you when you close the tab. The second generation will look like rooms. Pre-staged, inhabited, with named colleagues, a shared file system, and a memory that survives the session. The door matters.

My door took ten minutes to build. Yours might be a different file structure, a different launcher, a different mark on the icon. The pattern is portable.