Projects: the second axis of persona scope

Last week I wrote about backstory, when a persona needs to know who you are, and when it doesn't. This week I want to talk about the other axis. The one underneath.

A persona is the top-level container. Personal, Work, Family, the blogging persona, the dentist persona. But underneath each persona, real life has projects. Inside Family there's the kids' birthday party. Inside Work there's the Q3 launch. Inside Personal there's buying a house. These aren't tones, and they aren't tags. They're proper scopes in their own right, and they need to live somewhere.

The somewhere matters. If you don't get the nesting right, you get the same shared-context mush as before, just with a different sticker on it.

The shape

Here's how I think about it. The persona is the room. The project is the desk inside the room. The task you're doing right now is what's on the desk.

When I'm in the Family persona, working on the kids' birthday party project, doing the task of finalizing the invite list, the AI sees what's on the desk first, the desk second, and the room around it third. It does not see the desk in the next room over. It does not see the room down the hall. The Q3 launch over in the Work room? Doesn't exist as far as this conversation is concerned.

That nesting matters because it tells the AI what's in scope without me having to spell it out every time. And it tells the AI what's out of scope the same way.

Why projects shouldn't be a flat tag

The way most tools handle this today is with tags. Everything is in one shared context, and "project" is just a label you can filter by. The AI sees the whole pool but is told to weight the project-tagged items higher.

This is the wrong shape, and you can feel it the first time it breaks.

The Family persona is talking about the birthday party. Someone mentions "the venue." The AI, helpfully, pulls in a memory tagged "venue", except the memory is about the off-site venue for the Q3 launch from last year, because both items got the same tag at some point. The party planner now thinks the venue has a projector budget.

This is small, but it's the shape of the bigger problem. Tags don't isolate. They sort. And when the AI is allowed to see across the sort, it will.

Nesting projects inside personas changes the default. The Q3 launch isn't tagged "Work." It lives inside Work. There is no path from the birthday-party context to the Q3 launch context unless I, the user, explicitly take one. The AI doesn't have to figure out what's appropriate. The structure already did.

How projects shape what the AI volunteers

The other thing nesting fixes is the volunteer problem. AIs love to help. Left to their own devices, they'll fish out anything they think is even tangentially relevant. With flat tags, "tangentially relevant" is a wide net.

Inside a project, the net is the project. The AI working on the kids' birthday party should be volunteering: "you haven't confirmed the cake yet," "the guest list is at 14, the venue caps at 12, do you want to bump it?" It should not be volunteering: "you have a doctor's appointment that week," "by the way, the Q3 launch needs your sign-off." Those are real, but they belong to other scopes.

When the project is a proper container, the AI's helpfulness gets bounded. It can be aggressive inside the box and silent about everything outside it. That's the experience you want.

Why this matters for what the AI retrieves

The buzzword for "the AI pulling in relevant context" is RAG, retrieval augmented generation, the thing where the AI looks up information from somewhere and uses it to answer (this is called RAG, if you want to look it up later). Whatever you call it, the question is the same: when the AI needs to pull in extra context for a question, where does it look?

If the persona/project structure is a real container, the retrieval pool is the container. Inside the Family persona, the kids'-birthday-party project, the retrieval pool is the project's notes, the party's guest list, the messages in that thread. It is not the entire Family persona, and it is absolutely not everything the AI has ever known about me.

Same for memory. The AI's memory of "we decided on the rainbow theme" lives inside the birthday-party project. It does not need to be loaded when I'm doing taxes. It does not get loaded when I'm doing taxes. (I wrote more about that in the memory isolation piece coming up later in this series, for now, just hold the picture that memory is scoped where the work is happening.)

This is the part that actually makes the AI feel coherent. Not bigger models. Not longer context windows. Scope.

The three audiences, same nesting

Want the full container model first? Start with A persona is a container, not a costume.

Personal. I have a "buying a house" project inside my Personal persona. It holds the offer letters, the inspection reports, the back-and-forth with our agent, the running list of "what we still need to ask the seller." The Personal persona itself stays broader, fitness, reading, side ideas, the general me-outside-of-work. The house project is a sub-room. When we close, the project archives. The Personal persona keeps going. Importantly, the AI helping me with the house does not pull in my reading list or my running log just because they're also under Personal. They're not in this project.

Small Business. If you run a side business, the project axis is where the work actually lives. Inside the Business persona, you have a project per client. The "Acme Corp engagement" project holds the contracts, the meeting notes, the deliverables, the running tally of hours. The "Beta LLC engagement" project does the same for the other client. Both are inside the Business persona. Neither sees the other's notes. When the AI drafts a status email for Acme, it's working inside the Acme project, and a sentence about Beta's project couldn't sneak in if it tried, because Beta's project is in a different room down the hall.

Enterprise. At a real company, this is the difference between a real audit story and a vibes-based one. Inside the Work persona, there's the Q3 launch project, the SOC 2 prep project, the migration project. Each one has its own people, its own documents, its own AI memory. When an AI working inside the Q3 launch project drafts a status update for the launch team, the retrieval pool is the launch project. The SOC 2 prep project's documents aren't in the pool. The migration project's risk register isn't in the pool. If a question comes up that requires cross-project context. That's a deliberate decision, and it gets logged. (The fact that the AI itself has its own identity inside each project, and the audit trail names that identity, is the first-class-identity thread. I keep coming back to it for a reason.)

Same shape everywhere. The persona is the room. The project is the desk. Scope is what keeps the AI useful instead of overhelpful.

Projects without backstory

One of the things I love about the project axis: most projects don't need backstory. The persona above them has whatever backstory is appropriate (see last week's piece on that). The project just needs the project's facts.

The kids'-birthday-party project knows: who's invited, the venue, the theme, the budget, the cake decision. It doesn't need to know which kid this is for in a deep-history way. That's the Family persona's job to carry. It doesn't need to know my professional title, that's not in this room.

The Q3 launch project knows: the launch date, the feature list, the GTM owner, the open risks, the demo script. It doesn't need to know my career history. It doesn't need to know what the Q2 launch did, unless I explicitly say "pull in Q2 retrospective notes", which is a deliberate cross-project move, with a log entry attached.

This is the part that surprises people: projects feel lighter than personas. They have less to know. That's correct. The project is the most local scope. It should know the least and act the most.

The parts that will bite you

A few sharp edges to watch for.

Projects with no end date. Some projects are real projects, they ship, they close, they archive. Some "projects" are actually ongoing programs and they should be modeled differently. If your "project" has been open for 18 months and shows no sign of ending. It's not a project. It might be a sub-persona. It might be a recurring program. Decide which.

Shared documents across projects. Real life has overlap. The kids'-birthday-party project might reference the family's standing dietary-restrictions doc, which lives at the Family persona level. That's fine, projects can read up to their persona. But projects shouldn't read sideways into sibling projects without an explicit decision. The dietary-restrictions doc belongs to Family; the cake decision belongs to the party.

Cross-persona projects. Sometimes a project really does span personas. "Booking a personal trip" sits in Personal but touches the shared household calendar. "Refinancing a house" touches Personal and the joint-finances overlap a household maintains. These exist. Handle them with explicit cross-scope policies, not by collapsing everything into a flat tag pool. The whole point is that cross-scope is a decision, not a default.

Projects that should have been a task. Don't promote every task to a project. If the work fits in two messages and ends today. It's not a project. It's a task inside whatever scope you're already in. Projects deserve the overhead they carry; tasks don't.

What I'd do if I were setting this up

If you're standing this up (for yourself, for your business, or for your company) I'd start with three personas (Personal, Work, Family is a good first cut) and then write down the projects that are actually live in your life right now under each one. Not aspirational. Not "things I might do." The five-ish things you're actively moving on.

That list will be shorter than you think. That's good. Each one becomes a real container, with its own notes, its own memory, its own retrieval pool. The AI working inside it will feel sharper than the AI you had before, not because the model got better, but because you finally told it where it was.

Next week I want to talk about what actually changes at the model layer when the persona switches, the prompt, the tools, the memory, all at once. Because once you have personas and projects sitting properly nested, the question becomes: how does the AI know which room it's in? That's the next piece.