Project · Brand systems
Agent-Ready Brand Consistency Skills
I built two Claude Code skills — make-design and make-voice — that capture how a product looks and how it sounds, once, into two files: a DESIGN.md and a VOICE.md. Every agent that writes copy or builds UI reads them and checks its own work against them, so tone and design stay consistent across a project instead of drifting — and I spend far less of every session re-explaining the brand to Claude.
Two files — one for the look, one for the voice — that every agent reads before it writes.
Brand drifts when it lives in someone's head.
On its own, a project's identity lives in scattered places: a few posts that happen to capture the right tone, a mood board, some colours someone picked once, and a lot of it only in my head. So every time an agent writes a landing section or builds a component, it guesses, drifts a little, and I correct it again. The same notes about voice, the same fixes to colour and spacing, session after session. Adjectives like "friendly" or "clean" don't actually constrain what an agent produces; nothing is written down in a form it can be held to. The hard part isn't picking a palette or a tone — it's capturing both so an agent can load them, follow them, and check itself against them, and the brand stays consistent without me in the loop every time.
One captures the look. One captures the voice.
Each is a Claude Code skill that interviews me, then writes a single file at the project root. What's shown here is the shape and the output — the skills' own interviews and prompts stay inside them.
A structured interview about positioning, mood, and surfaces, then a deterministic pass that turns the answers into a spec-conforming design system — not a mood board, a file a UI can be built from.
- Colour tokens, validated for contrast
- A type scale and spacing rhythm
- Shape and component definitions
- Exportable to the formats UI is built from
An interview about who's speaking and to whom, then a calibration step that writes real sample passages I react to until they land — because examples constrain an agent, and adjectives don't.
- Gold-standard sample passages
- A banned-word list
- Tone, register and mechanics
- A self-check the agent runs
Interview, validate, lock — then everyone reads it.
Both skills run the same shape. It's a gated pipeline, not one-shot generation: I'm asked the questions, the answers get validated or calibrated, and only then does the file get written — so what lands is something I've actually signed off on.
The files aren't documentation — they're inputs.
A style guide nobody loads still drifts. These two files are read by the agents doing the work, right when they do it — so the brand is enforced where it's used, not just written down.
Any skill or agent writing copy reads VOICE.md and checks its draft against it before shipping — so a landing section or an error message sounds like the brand, not like a generic model.
Anything building an interface pulls tokens and rules from DESIGN.md instead of re-picking colours and spacing — so screens stay visually consistent across a project.
The same file I open to understand the brand is the file an agent loads to conform to it. There's no second copy to drift from.
When the look or the voice changes, I re-run the skill on the part that changed — so the spec stays coherent instead of accumulating brittle manual edits.
Design tokens are checked for contrast; the voice is calibrated against passages I signed off on. The spec is something approved, not a first draft.
Because every token carries a value and every voice trait carries an example, an agent can execute the spec and verify itself — the part that makes hands-off automation safe.
Author the brand once; every agent after it conforms.
The work of deciding how a product looks and sounds gets done a single time, in a single place, and validated before it's locked. After that, staying on-brand isn't a note I repeat every session — it's a file every agent already reads and checks itself against. Less back-and-forth, more autonomy, and automation I can actually trust to ship.
On-brand without me re-checking every post.
The same VOICE.md and DESIGN.md are what keep a one-draft-to-fifty-posts pipeline on-brand across every surface — no human re-checking the tone of each post.
Learn more→Point a fresh project at its own DESIGN.md and VOICE.md and every agent starts out already knowing how it should look and sound.
The decisions I keep are the ones worth keeping — the brand itself. The repetition of enforcing it, session after session, is gone.
Want your brand to run itself this cleanly?
Tell me what you're building, and I'll come back with whether I can help and what a first step looks like.