A skill for Claude Code that designs a brand identity as one system: **logo, typography and colour palette built
together from one idea**, measured, shown side by side as identity cards, and handed over as a brand guidelines kit
for the set you pick.
Twenty-four briefs, one direction shown from each: a light festival in Lyon, a taquería in Mexico City, a
pharmacy in Athens, a techno label in Berlin, a floating sauna in Oslo, a children's dentist, a weeding-robot
start-up, a honey co-op on the Black Sea and more. Mascots, emblems, monograms, stencils and plain wordmarks; every
mark, typeface and colour above came out of a run of this skill.
**[What you get](#what-you-get)** · **[Examples](#examples)** · **[Install](#install)** · **[Use](#use)** ·
**[How it works](#how-it-works)** · **[Network and privacy](#network-and-privacy)** · **[Develop](#develop)**
## What you get
- **Three (up to four) identity sets**, each built around a single mechanism that produces the mark, the type choice
and the palette at once, so the parts belong together instead of being picked separately.
- **Measured, not asserted.** Contrast (WCAG 2.2), colour-blindness simulation, glyph coverage for the
languages the brand writes in, font licences, tabular figures, and logo renders at 16 and 32 px for review. A failed
measurement is a gate the set must pass; a heuristic is labelled as a warning; taste is labelled as judgement.
- **Sets that don't look like every other AI brand.** The skill knows what plain models reach for and asks for a
reason from the brief whenever a set uses it ([below](#why-the-sets-dont-look-like-every-other-ai-brand)).
- **A brand guidelines kit** for the chosen set: an 8-page PDF (cover, idea, logo, variations, logo on colour and
misuse drawn for *this* mark, colours, typography, applications; `--full` adds grid, dark mode, accessibility,
iconography and imagery for 13), logo files (SVG and PNG: full colour, one colour, reversed, app icon, favicon), design tokens (CSS,
SCSS, Tailwind, DTCG JSON) and a font sheet with embed code.
- **Your site, restyled.** Give it a URL and it reads the current logo, type and colours, then previews each set on
your own pages. Keep the parts you like (`keep`), evolve others (`refresh`).
- **Any part on its own**: just a logo, just a type system, a palette for an existing logo, or a critique of an
identity you already have.
## Examples
### Fernhill Books: an independent bookshop and small press in Edinburgh
*Brief: literate, unhurried, wry. Competitors: a national chain in navy and white, a second-hand shop with a gilt
sign, an online marketplace.*
**A · Manicule** (recommended, picked): the pointing hand readers drew in margins is the mark, Literata carries both
display and text, and the palette is press black with one rubric red. **B · Full Stop**: a quiet lowercase name that
ends on a blue proofreader's full stop. **C · Close**: an Old Town passage under a tenement, on a dark ground.
| Kit: cover | Kit: typography |
|---|---|
|  |  |
### Tessellate: an open-source data pipeline orchestrator
*Brief: precise, calm, candid; the product shows run times and costs, so figures must be tabular. Competitors: an
established scheduler with a pinwheel logo, two start-ups in purple and teal.*
**A · Truchet Tile** (recommended, picked): one tile cut by two quarter arcs that join any neighbour however it is
turned, the way steps of a pipeline meet. **B · Readable Run**: a text serif wordmark whose first *t* carries a
duration bar. **C · Night Floor**: two L-pieces lock into one square, dark-first.
| Kit: logo on colour, and misuse specific to this mark | Kit: applications |
|---|---|
|  |  |
### More runs
| Brief | Languages | Directions (recommended first) |
|---|---|---|
|  **Molino Coffee**, a specialty coffee shop in Lisbon | en, pt | Vizinho · Mó · Bloco |
|  **Hartwell & Oduya**, a commercial law firm in London | en | The Bar Between · Equal Measure · Counterparts |
|  **Parallax Labs**, AI planning agents for supply chains | en | Depth Field · Parallax Error · Computation Pad |
|  **Domaine Fauvel**, a natural wine estate in Burgundy | en, fr | La Taille · Climat · Fauve |
|  **Moonvault**, a self-custody crypto wallet | en | Held Moon · Two O's · Lamp Arcade |
|  **Hasat**, a Turkish breakfast restaurant in Istanbul | tr, en | Tepeleme · Demli · Uzun Sofra |
## Install
**Claude Code plugin** (recommended):
```
/plugin marketplace add fatihaydost/brand-identity-skill
/plugin install brand-identity@brand-identity-skill
```
**Or as a plain skill**, from the `brand-identity.zip` of the
[latest release](https://github.com/fatihaydost/brand-identity-skill/releases/latest) (unzip it into
`~/.claude/skills/`), or from a clone:
```bash
git clone https://github.com/fatihaydost/brand-identity-skill
mkdir -p ~/.claude/skills
ln -s "$PWD/brand-identity-skill/skills/brand-identity" ~/.claude/skills/brand-identity
```
**Requirements:** Python 3.10+, a Chromium-based browser (Chrome, Chromium, Edge or Brave) and, ideally,
[uv](https://docs.astral.sh/uv/). With uv the scripts install their own Python packages (fonttools, uharfbuzz,
skia-pathops) into a cached environment on first use; nothing to `pip install`. Run
`python3 skills/brand-identity/scripts/brand.py check` to see what is missing. Tested end to end on Linux, including
a clean container; unit tests run in CI on Linux, macOS and Windows.
**Other hosts.** The skill is a standard `SKILL.md` folder, so agents that read Agent Skills can load
`skills/brand-identity`; it is tested end to end in Claude Code. In claude.ai chat the code sandbox does not allow
Google Fonts by default, so type measurement and rendering fail there; use Claude Code.
## Use
Ask Claude Code in plain words:
- "Logo, fonts and colours for a specialty coffee roaster in Lisbon."
- "New type and palette for https://example.com, keep the logo."
- "Just a type system for my newsletter."
- "Critique our brand identity."
It asks one round of questions (what to make, how many sets, languages, competitors), builds the sets, shows the
board and **stops** for you to pick. It never picks for you. Then it builds the kit for your choice, or for a mix
("logo from A, palette from B").
A full run to the board takes about 40 agent turns and roughly 0.9 M billed-input-equivalent tokens (measured on six
briefs; see [testing](docs/testing.md#budget)).
## Why the sets don't look like every other AI brand
Asked for a brand identity, plain models converge: cream ground, terracotta, a gold accent, Fraunces headings, Inter
body text, a logo that hides a letter in a circle. The skill knows these defaults and asks for a reason from the
brief whenever a set uses one; at least one set per run uses none. On the same six briefs:
| | AI-default hits (type and palette) | sets with none |
|---|---|---|
| Plain model, 24 answers | 4.38 per answer | 0 of 24 |
| brand-identity, 18 sets | 0.28 per set | 13 of 18 |
Method and data: [`docs/testing.md`](docs/testing.md#ai-defaults).
## How it works
```mermaid
flowchart LR
Q[One round of questions] --> S[sets.json mechanism, mark, type, palette seeds]
S --> L[logos contact sheet at 16 and 32 px]
L -->|fix one defect per mark, twice| S
S --> B[build palette roles, audits, cards, board]
B -->|a gate failed| S
B --> P{You pick}
P --> K[kit PDF, logo files, tokens, font sheet]
```
The agent designs; the scripts measure. Gates (contrast, licence, glyph coverage, file validity, renders) stop a set
until it is fixed; warnings are heuristics and say so. Architecture: [`docs/architecture.md`](docs/architecture.md).
Tests (unit, smoke, rendered-output probes, end-to-end runs): [`docs/testing.md`](docs/testing.md).
## Network and privacy
The skill runs locally. It contacts:
- **Google Fonts** (`fonts.googleapis.com`, `fonts.gstatic.com`) and the
[google/fonts](https://github.com/google/fonts) repository on `raw.githubusercontent.com`, to download fonts and
their metadata and licences for measurement and rendering;
- **PyPI**, through uv, once, for its Python packages;
- **your site and the competitor sites you name**, opened in a headless browser, only when you give their URLs.
No telemetry; nothing is uploaded. The brand's fonts are never subset or redistributed: the kit links to them and
gives embed code. Licences are checked per family (OFL by default; Fontshare and commercial faces get a licence note).
**Your brief and Claude.** The skill sends nothing anywhere, but what you tell Claude (the brief, the strategy, client
and competitor names) is processed by Anthropic under your plan's terms. On Free, Pro and Max, including Claude Code,
retention and training follow your privacy setting
([consumer terms](https://www.anthropic.com/news/updates-to-our-consumer-terms)); under the Commercial Terms (Team,
Enterprise, API) Anthropic may not train models on customer content
([Commercial Terms](https://www.anthropic.com/legal/commercial-terms)). Check your plan before pasting confidential
client material.
**Trademark clearance is yours.** The skill does not search trademark registers, and Anthropic's IP indemnity under
the Commercial Terms excludes trademark claims based on using an output in trade or commerce. Before a logo goes into
use, have it cleared in the countries where the brand will trade.
## Develop
```bash
pip install -r requirements-dev.txt # requirements.txt + gflanguages (checks the bundled language data)
python3 -m unittest discover -s tests
python3 tools/smoke_test.py
python3 tools/package_skill.py # dist/brand-identity.zip
```
```
skills/brand-identity/ the skill: SKILL.md, references/, scripts/, templates/, assets/
tools/ smoke test, packager, run measurement, AI-defaults experiment, catalogue builders
tests/ unit tests
evals/ behaviour evals and fixtures
docs/ architecture, testing, images
```
## Licence
MIT, see [LICENSE](LICENSE). Adapted code and bundled data: [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md). Brand
names and measured site colours belong to their owners: [TRADEMARKS.md](TRADEMARKS.md). The example briefs in this
README were invented for it; any likeness to a real business is coincidental.
