Brand memory that grows while you design.
You already have taste — it shows up every time you tell an AI "not this one", "warmer", "less rounded", "yes, that's it." Then the session ends and all of it evaporates. Next session, the AI starts from zero and you correct the same things again.
growing-brand is a thin layer for AI-assisted design work (built for Claude Code; it's all plain markdown, so any agent that reads skills can use it). It rides along inside your normal work, catches those judgments as they happen, and grows them into a brand system you never had to sit down and write.
No routine. No ratings. Nothing to run. You just work — and the work stops forgetting.
Three things happen automatically, inside whatever you were already doing:
- Apply — before making design choices, the AI loads your
brand/files and wears them. The more you've worked, the less generic everything gets. - Capture — when you pick, reject, correct, approve, or declare a rule, that judgment lands as one line in
brand/journal.md, in your own words:7. 2026-07-05 [corrected] buttons rounded-full → 4px — "we're not a toy" (job: pricing page). - Distill — when enough judgments pile up (~10), they're automatically summarized upward. Repeated judgments become rules at the right level; every rule cites the journal entries that earned it. You get a two-line note about what was promoted — and pushing back on a promotion is itself a judgment.
Four plain markdown files in your repo — yours to read, edit, or delete. Hand edits are treated as the strongest signal you can give.
| File | Holds | Changes |
|---|---|---|
brand/core.md |
Identity — what we always do, what we refuse | rarely — only rules proven ≥3 times across jobs |
brand/tokens.md |
The measurables — color, type, spacing, shape | sometimes |
brand/moves.md |
Recurring patterns — how things tend to be done here | as habits emerge |
brand/journal.md |
Every judgment, raw and numbered | every session |
Contradictory evidence is never resolved by fiat: it either splits by context ("4px on product UI, 12px on marketing") or sits openly in core's Tensions section until more evidence settles it. Rules can be demoted too.
See example/ for a fictional brand after three jobs — including which rules got promoted and why.
As a Claude Code plugin:
claude plugin marketplace add vivian-chi/growing-brand
claude plugin install growing-brand@growing-brand
(Or from a local clone: claude plugin marketplace add /path/to/growing-brand, then the same install.)
Or just copy the skills (no plugin system needed):
cp -r skills/* your-project/.claude/skills/ # this project only
cp -r skills/* ~/.claude/skills/ # every project (one brand/ per project)
For the most reliable auto-triggering, add one line to your project's CLAUDE.md:
Use the brand-memory skill for all design and visual work.
Run /brand-init in your project. If you already have design work there (CSS, tokens, components, a live site), it seeds the measurables from what exists — marked unconfirmed until real use proves them. It asks at most three one-line questions about identity; all skippable. Cold start is fine — empty files plus the journal will grow everything.
Or don't even init: the first time you make a design judgment, the layer offers once to start remembering. Say yes and keep working.
Where does my data go? Nowhere. Everything is local markdown in your repo. Nothing is sent, synced, or stored anywhere else.
Can I edit the files by hand? Please do — they're yours. A hand edit counts as the strongest kind of judgment.
Does it generate designs? No. It's a memory layer that sits under whatever generates — it makes the output wear your accumulated taste instead of the model's defaults, and it gets sharper every time you react.
Teams? v0 is one brand per project, grown by whoever works in it. Shared/team flows are on the roadmap.
Why files and not a database? So the brand is portable, diffable, reviewable, and still yours if you stop using the tool.
v0.1 — young and honest about it. The loop (apply → capture → distill) is fully specified and tested end-to-end; the file format may still evolve. Issues and picks-over-flattery feedback welcome.
MIT © Vivian Chi