thehonestape/brand-voice-spec

Brand Voice Spec

A machine-readable format for steering an LLM toward a specific brand voice, with a complete worked example. The point is not the example brand. The point is the method: treat brand voice as data a model can load and enforce, and as a living artifact that learns from its own corrections.

Most brand voice lives in a slide deck no model can read. When an LLM writes copy, it falls back to the median of its training data: hedging, buzzwords, passive voice, the corporate gray. This spec is the inverse. The same file a human maintains is the file that conditions the model.

The worked example is a fictional brand, Lantern Coffee, an independent roaster. It is invented. Fork it, delete the contents, and write your own.

What's in here

The canonical artifact is voice.json. SCHEMA.md documents the format. Everything under data/ is derived from voice.json and loads in the dataset viewer.

Config	Rows	What it is
rules	8	The voice rules. Each has a principle and, where it exists, the id of the correction that produced it.
examples	28	On-voice and off-voice copy, each labeled (on_voice / off_voice), tagged by surface and register. These are independent pools, not 1:1 pairs.
preference_pairs	5	Genuinely matched chosen / rejected lines. DPO-shaped. Use these for preference work; do not synthesize pairs from examples.
lexicon	93	Banned terms, preferred terms, domain vocabulary, and signature phrases.
prompts	4	Parameterized generation prompts, one per recurring writing job.
changelog	4	The voice's evolution. Each revision links to the correction that caused it.

Two ideas worth stealing

Registers. One brand can hold more than one voice. Lantern has a cafe register (guest-facing, warm, plain) and a trade register (wholesale, precise, price-transparent). Same brand, different surface, different voice. Every example and prompt is tagged with the register it belongs to.

The loop. Voice rots when it's written once and frozen. Here, a wrong draft becomes a note in voiceQueue, a periodic sweep promotes adopted notes into guidelines, the revision bumps, and the changelog records what changed and why. changelog joins back to the originating correction, so you can read the whole voice as cause and effect:

{"revision": 2, "change": "Added 'Talk like a regular, not a sommelier' after repeated sommelier-speak in menu drafts.",
 "derived_from_rule": "Ban tasting poetry. Concrete anchors only.",
 "derived_from_why": "'Notes of bergamot dance across the palate' kept reappearing in menu copy."}

The rule is not an opinion someone had in a meeting. It is the scar from a specific mistake.

Load it

from datasets import load_dataset

rules = load_dataset("thehonestape/brand-voice-spec", "rules")
pairs = load_dataset("thehonestape/brand-voice-spec", "preference_pairs")

Uses

• Few-shot conditioning. Drop rules + bannedTerms + a handful of preference_pairs into a system prompt.
• Preference tuning. The pairs are ready for DPO or reward-model experiments on style.
• Voice-lint evaluation. Score generated copy against bannedTerms and the rules.
• A template. Fork voice.json, write your own brand, regenerate the data.

Provenance and license

Built by Abe Garcia at Workhorse, an independent design studio. The format is the one we use in production to keep AI-written copy on-brand across every surface. Lantern Coffee is fictional. Released under CC BY 4.0.