Russian Dolls: A Better Way with Claude Cowork
Show a child a Russian doll “Babushka” and watch their delight as they open up one by one getting smaller and smaller. Each doll is more specific, more concentrated as the outer shell gives way to the inner details, and you can’t reach that detail without passing through the shell first. They’re also very cute. I have one on my shelf as I write.
That image is the best explanation I have for how I now work with Claude. It came courtesy of Jeff Su’s CoWork Tool Kit which introduced me to what he (and now I) call the stacking system: a layered workspace architecture that Claude reads before you’ve said a single word. Trust me, this one works. I’ve been using it now for a couple of weeks and it’s saved me time, effort, and credits!
The problem it solves is a familiar one. You open a new session in Claude CoWork and the thing that was your sharp, context-aware collaborator yesterday is, today, a stranger. You re-explain your tone. You re-establish your projects. You remind it, again, that you use British spelling and prefer a direct recommendation over three hedged options. The stacking system (built on Claude’s ability to read layered instruction files) puts an end to that.
How it works
Inside a dedicated folder on your computer (or in the cloud) and at the outermost layer sits a CLAUDE.md file at the root of your workspace. Here are your universal rules, your non-negotiables: how you want things written, named, and structured. Alongside it, there’s a MEMORY.md that holds the facts that change over time: active projects, contacts, decisions made. These two files are the shell that wraps everything.
Open that outer doll and you find your individual workstations: folders for the distinct domains of your work and life. For me that’s work, finances, travel, health and fitness. Each has its own CLAUDE.md (rules specific to that domain) and its own MEMORY.md (context relevant only there). A routing map in the root tells Claude which workstation to load based on what you’re asking about. You don’t navigate manually. The system does it.
Open a workstation and you find projects, each inheriting everything from the layers above and each in their own folder. They also have their own CLAUDE.md and MEMORY.md files. Reference files for a workstation or a project sit in their own resources folder, and are pulled in only when a trigger condition is met. The whole thing reads outer to inner before Claude responds to anything. The most specific context always applies, and the global rules always hold. See how the stacking works? Neat, eh?
Why it works
The first benefit is the obvious one: you stop repeating yourself and so you save time. Every preference, every convention, every piece of context you’ve established is already loaded. The second is subtler. Because MEMORY.md accumulates over time, Claude’s responses get more accurate and more tailored the longer you work this way. There is a real difference between a contractor who arrives cold and a colleague who has been working alongside you for months.
The third benefit is focus. Claude gets smart enough in time to know which Workstation to go to and which Project. When you start working, just tell it what you want to do and away you go. Claude is ready. Because context is scoped to the active workstation, Claude isn’t carrying noise from domains you aren’t currently working in. When you’re reviewing finances, Writing isn’t in the room, so your sessions stay cleaner and leaner.
The fourth is practical and, once you notice it, hard to un-notice: the system is easier on your credit allocation. Lean context means smaller processing windows, which means slower credit burn. If you work across multiple domains in a day as I do, you’re running focused sessions rather than one bloated context hauling everything at once. Over weeks, that adds up.
And then there’s the benefit that matters most in the long run: the intelligence lives in files you own: plain text, version-controlled, portable. If the tool changes, the model changes, or you decide to move somewhere else entirely, your context comes with you. You are not locked into anyone’s memory system but your own.
The discipline it requires
None of this works without maintenance. The distinction that holds the system together is simple: CLAUDE.md is for behaviour (how things should be done), MEMORY.md is for facts (what is currently true). Blur that line and the files get noisy. Jeff (and I) archive completed projects. This keeps entries lean. The system pays you back in direct proportion to how tidy you keep it.
What Jeff Su’s Cowork Toolkit gave me was not a new tool but a new way of thinking. Each layer of the workspace narrows the scope and adds precision. By the time you type your first message, the work of establishing context is already done. Check it out for yourself.
Thanks, Jeff.
Photo by Julia Kadel on Unsplash