Last week a colleague asked me how I set up the folder for my personal brand. Not the site, not the newsletter. The folder. The thing on my hard drive that everything else came out of.
I started answering and realized I’d never written it down, which is funny, because writing things down is the entire method.
Then, two days later, I went to set up the same kind of folder for a new company I’m building with a partner. I ran it against my personal one, layer by layer. It failed four of the seven checks. The company has the best decision trail of anything I’ve built: every rejected name, every killed direction, dated. A fresh session of Claude Code would still have opened the wrong file and built on a corpse.
So here’s the anatomy. Seven layers, what each one is for, and the test that told me where the new folder was thin.
Why a folder and not a prompt
I’ve made the argument before, in The Company Is the Context, so I’ll keep it short: the model doesn’t need a better instruction. It needs the world it’s working in. For a person or a brand, that world is who you are, how you sound, what’s been decided, what’s current, and what you’ve corrected. Those are five different kinds of knowledge and they change at five different speeds. A prompt jams them into one paragraph. A folder gives each one a home.
The rule I’ve landed on: every layer has one obvious home and one source of truth. When something changes, I change it in one place and the rest follows. When a session starts, it reads the layers in order and arrives already knowing the person.
That’s it. Everything below is that rule applied seven times.
Layer 1 — The foundation, above the brand
I have two brands: my personal one, and the agency. For a while I tried to write two identities. They kept drifting apart in ways that felt wrong, because they come from the same person and the same idea.
So I wrote one file that sits above both. It holds the origin (both trace back to Joseph Campbell’s Hero’s Journey, which is in my name and in the agency’s), three beliefs, a personality, and a design sensibility. Both brands inherit it. Neither dictates the other. The new company inherits it too, which is how I know the three will feel related when they sit side by side.
A note for the record: I have no affiliation with the Joseph Campbell Foundation. Campbell’s work is an inspiration to me, nothing more, and none of this is endorsed by anyone but me.
The order I’m working in matters. I started with my personal brand, and that update is done. Now I’m building and launching the new venture and its brand, which is what this week’s audit was for. Then the Campbell Learn brand gets revised against both. Each one inherits the foundation, and each one teaches the foundation something on the way through.
The important part is the order. The foundation was written first, before any palette, before any tagline. When I open a session in any of the three folders, that file is the first thing the model reads. It’s the file I’d hand a new hire on day one if they asked “what is this all actually about.”
Layer 2 — The operating file at the root
Claude Code reads one file automatically when it starts in a folder. Mine for the personal brand is short: who this is, the voice rules that can’t be broken, and where everything lives.
The voice rules are the part that earned its keep. They came out of arguments I lost with myself. I rejected “a writer who builds.” Then I rejected “a builder who writes.” Then I understood the problem: any “X who Ys” makes one identity the noun and the other a hobby. The containing noun is human — writer, builder, founder, father, all at once. That rule is one line in the operating file, and every piece since has honored it without my having to say it again.
The other rule that lives there: AI-native optimism, never guarded. The era is an invitation. Any draft that treats the tools as the villain gets killed at the operating-file level, before I ever see it.
This is where the new company’s folder failed first. Its operating file was a naming rule and a directory map — accurate, useful, and silent on who the company is. The identity existed, three files deep. Nothing told the model to read it first.
Layer 3 — Strategy the model can cite, with the rejects
Under the brand folder there’s a positioning doc, an ecosystem doc, and a running checklist. What makes them work isn’t the decisions. It’s the rejected options, dated, next to them.
The positioning doc records that “returning writer” was retired as public framing, and why (“I don’t get the returning writer part”). It records that the tagline moved from an imperative, keep the human in the loop, to an identity statement, the human in the loop, and the date that happened. When the model drafts a bio, it isn’t guessing my taste. It’s reading the transcript of how the taste got decided.
The new company does this better than I do, honestly. It keeps three tiers: a thinking log for ideas before they harden, a foundation doc for drafted positions, and a decision record for what’s closed. If you’re starting from scratch, steal that shape.
Layer 4 — The design system, plus one file that says what’s current
The palette lives in a single Brand DNA file. Every hex value was sampled from one source illustration: a father and child under a starry navy cloud, holding a cursor. Nothing invented. The file explains why the palette works (one warm family, one cool mass, one gold spark) and gives usage ratios and rules. That’s the spec.
Next to it is a file called CANONICAL, and it answers exactly one question: what is current, right now. A table of roles (hero, mark, scene library, palette) and for each, the file to grab and the master it points to. The first line says: if a filename and this file disagree, this file wins.
I wrote that file after a bad week. The hero illustration was locked and right. Then I had the model generate character sheets, turnarounds and poses and expressions, to extend the world. They came back with a different man: longer hair, a full beard, a mauve jacket, clean vector fills instead of pen-and-ink. And then, because the sheets existed, a later session wrote that man down as the standard and generated four more sheets to match him. A made-up person had become canon by being the most recent file in the folder.
The lesson is in the CANONICAL file now, in bold: anchoring to a derivative instead of the source. The old design-system doc is still in the folder, with SUPERSEDED stamped at the top and the date. The rejected sheets are still there too, marked off-model. Nothing gets deleted, because the trail is the point. But nothing unmarked gets to pretend it’s current.
This is where the new company’s folder failed hardest. Fifty lab files side by side. A direction that was killed a month ago (“too similar to a competitor”) sitting there unstamped, one open-file away from being built on again.
Layer 5 — Memory, one fact per file
Claude Code keeps a memory directory per project, outside the repo. Mine for the personal brand has about sixty entries. They split into two kinds.
Project memories are facts: the newsletter runs on Kit, the illustration style is locked, here’s where the analytics live. Feedback memories are corrections: no likeness of me in illustrations; mockup before any production deploy; the site says “essays,” not “letters.” Each feedback entry carries a why and a how to apply.
The feedback kind is what compounds. Every correction I give once becomes a rule the next session already knows. I’ve watched the same note, the one about the site saying essays, go from something I typed in frustration to something the model catches on its own before I see the draft.
The failure mode here is subtle. The new company’s memory existed, but as one long file with nine dated updates appended to the bottom. Real rules were buried in paragraph six. “Machine parts, not pictures,” for one, which is why a whole vein of names got killed. A rule inside a scroll doesn’t get recalled. A rule in its own file does.
Layer 6 — Content streams that carry their own rules
Each kind of content I make has its own operating doc. The interview series I run has two. One covers what and where: the six outputs, the publish steps, the page layout. The other, called CRAFT, covers how: six to nine hundred words, open on a scene, at least three direct quotes, close forward, the subject is the hero, the byline is always mine.
The LinkedIn library has a catalog: every post, its date, its status, its topics, and what it can be repurposed into. Twenty-one posts in there now. When I ask for a new one, the model isn’t working from a blank page. It has the last twenty and the rules they followed.
The new company has no content stream yet, so this layer isn’t a failure yet. It’s a date. The day the first stream starts is the day its CRAFT doc gets written. Not after three pieces have drifted.
Layer 7 — Skills for the repeatable moves
A few things happen often enough that they became skills: small instruction files the model can call. Cataloging a post. Building a promo kit for a new piece. Resending a newsletter issue to the people who didn’t open it. Four of them live in the brand folder.
The test for whether something should be a skill: have I explained it twice? Then it’s a skill. A memory can be skipped on a given day, because memories load by relevance. A skill loads when it’s called. I wrote about this distinction in The Creative Unlock — the taste that stays in your head doesn’t exist next session.
The audit — run it on your own folder
Here’s the test I ran on the new company. Seven questions, one per layer. Answer each from one file, or you’ve found the gap.
- Is there a foundation above the brand? Something that would still be true if the brand changed its name.
- Does the root file say who this is, in a paragraph, with the voice rules that can’t be broken, before it says where anything lives?
- Can the model cite a rejected option? What was decided, and also what lost, and why, and when.
- Is there one file that says what’s current right now? If a filename and that file disagree, which wins?
- Is memory one fact per file, with a why and a how to apply on every correction?
- Does each content stream have a craft doc with word counts and rules a stranger could follow?
- Has anything been explained twice that should be a skill?
The new company passed 3 and 6 (the second one by not having content yet). Failed 2, 4, 5, and 7. I spent an hour fixing it: rewrote the operating file to lead with identity, added the CANONICAL file, stamped the dead labs, split the memory. An hour, because the pattern already existed one folder over.
What the folder is actually for
I keep calling it a context system, and it is. But the honest description is smaller and stranger. The folder is where my taste goes so it can exist without me in the room.
Every argument I’ve had with myself about what I am. Every hex value I sampled from a drawing. Every time I said “not that, this.” It’s in there, in the one place the next session will look. The design system came fast once the folder held the person. The content came fast once the folder held the rules. When I tried it in the other order — assets first, identity later — it came slowly, and wrong, and I couldn’t say why.
Write the foundation. Write the operating file. Then build.
Joseph Campbell’s Hero’s Journey is a personal inspiration. I am not affiliated with, sponsored by, or endorsed by the Joseph Campbell Foundation.
— Joe
