Running several projects through one AI setup is easy. Keeping them from sounding like each other is the actual work.
The answer is a hierarchy rather than separation. One shared layer holds everything that must never vary, and each project holds only what makes it itself:
Its language, its market, its tone, its palette, its destination. Nothing gets duplicated, and nothing bleeds, as long as you’re deliberate about which level a rule lives on. This is one layer of a larger setup, and it’s the one that decides whether the rest of it scales past a single brand.
Two levels, and only one direction
The shared layer sits on top. Each project sits underneath it and extends it. That word is doing real work: a project file adds and specialises, and where the two ever disagree, the shared rule wins.
That direction matters more than it looks. Let a project override the shared layer and you’ve built five independent systems that happen to share a folder, which is the situation you were trying to avoid. Keep the direction one-way and the shared layer stays genuinely load-bearing, so a rule written once is enforced in every project including the ones you build next year.
The direction also settles what to do with something you learn inside one project. If the lesson would hold everywhere, it gets promoted upward rather than copied sideways. That one habit is what stops five projects becoming five slightly different versions of the same rule, each drifting on its own schedule.
Most tools that read instructions from disk already implement this shape. Claude Code loads from broadest scope to most specific, so a user-level file loads before a project file and the closest file is read last. You don’t have to build the hierarchy. You have to decide what goes where.
The test for what belongs where
One question settles almost every case: would breaking this rule be wrong in every project you run?
If yes, it’s shared. Never fabricate an experience or a statistic. Never publish without a human reading it. The punctuation conventions. The banned constructions that give AI prose away. None of that gets better or worse depending on which brand it’s for, so none of it should be written down more than once.
If the answer is no, it’s a project rule. Language and market. Whether the tone is commercial or editorial. How hard the call to action pushes, which for me differs sharply between the project that sells and the ones that publish. Reading level. Sentence rhythm. What the pieces are for.
The cases that feel ambiguous usually resolve by asking what happens when you add a sixth project. If the new one would need its own version of the rule, it’s a project rule. If you’d copy the existing one unchanged, it was shared all along and you’d just written it in the wrong place.
What each project needs of its own
Four or five small files, and the list is stable enough that adding a project is mostly copying.
Who it’s for and what it’s against. Audience, positioning, editorial do and don’t. This is the one that determines whether the output has a point of view.
How it sounds. Language, rhythm, pronoun stance, the specific taboo phrases for this brand. Covered properly in the piece on why AI content sounds generic.
A structured version of the same thing. The persona as data rather than prose, so a checker can enforce sentence length and readability rather than a human eyeballing it.
What it looks like. Palette, composition, subject matter, and what to avoid. This is the one people forget, and it’s the one that gives brand bleed away fastest. Prose voice is subtle. Two projects using the same stock-photo aesthetic is visible from across a room.
Where it publishes. Each project points at its own destination, configured per project rather than chosen at runtime. A misrouted draft is a small disaster with a large recovery cost.
What doesn’t get its own version is anything already covered above it. The temptation early on is to hand each project a full copy of everything so it feels self-contained. It looks tidy and it’s the start of the problem: five copies of a rule means five places to update, and in practice one gets updated and four go stale. All of it lives in my own folders rather than inside any tool, which is what stops an update quietly undoing it.
Where the bleed actually comes from
Here’s the part that surprised me. Brands don’t blend because the tool confuses two sets of files. That failure is loud, obvious, and rare.
They blend at the points where no rule was ever written. Anything you haven’t specified falls back to the model’s default, and the default is identical for every project you run, because it’s the same model with the same average. So the projects converge precisely on the dimensions you were least deliberate about.
Which means the bleed is diagnostic. When two projects start sounding alike, the useful question isn’t which file is wrong. It’s what those two pieces have in common that neither file mentions. Usually it’s something nobody would think to write down: how a piece opens, how long the sections run, the transitional phrases, the way an argument gets set up before it’s made.
Write that dimension down, at whichever level it belongs to, and the convergence stops on the next draft. That’s the whole loop, and it gets shorter every time you run it.
It hasn’t actually bitten me yet, and I don’t think that’s luck. Two things keep catching it early. I’ve been watching for it, so I’ve stopped and corrected things before they had a chance to converge. And each project has a clear enough direction that a piece coming back in the wrong framing is obvious rather than subtle. Not slightly off. Visibly not what that project is for.
Which is worth saying, because the direction is doing work the files can’t. A written rule catches what you thought to write down. Knowing what a project is actually for catches the rest.
Adding the next project should be boring
The test of whether the layer is factored properly is what happens when you add one. Copy the file set from the nearest existing project, change what’s genuinely different, register it, done. If it takes an afternoon of thinking, the shared layer is holding something that should have been project-level.
Mine is six steps and none of them are interesting, which is exactly what I want from it. The interesting work is the positioning of the new project, and that’s a decision, not a configuration.
There’s a second effect that only shows up once you’ve done it twice. When setup is expensive, ending something feels expensive too, and you keep projects alive that should have been closed months ago. Cheap setup makes closing cheap. For a one-person operation, being able to close things is worth considerably more than being able to start them.
The brands never blend where the rules disagree. They blend in the places where nobody wrote a rule at all.
Build Your Content Machine. A free 5-part email course, one email per layer. Day 3 is this one, with the file set laid out. Start the free course →
