SCHEMA.md

# Brand Voice Spec — schema

`voice.json` is the canonical artifact. Everything in `data/` is derived from it. The format is small on purpose: enough structure for a model to load and enforce, not so much that a human won't maintain it.

## Top level

| Field | Type | What it is |
| --- | --- | --- |
| `brand` | string | The brand this voice belongs to. |
| `status` | string | `sample` for this reference; `live` for a real brand. |
| `revision` | int | Bumped every time a rule changes. Pairs with `changelog`. |
| `registers` | object | Named voices within one brand. Here: `cafe` (guest-facing) and `trade` (wholesale). A field can carry a `register` to say which voice it belongs to. |
| `tone` | string[] | Adjectives. The 30,000-foot read. |
| `guidelines` | object[] | The rules. See below. |
| `doAndDont` | object[] | Genuinely paired on-voice / off-voice examples. |
| `surfaceDoDont` | object[] | Per-surface guidance plus independent pools of good and bad examples. |
| `promptKit` | object[] | Reusable generation prompts, one per recurring writing job. |
| `bannedTerms` | string[] | Hard prohibitions. |
| `preferredTerms` | string[] | Reach-for-these words. |
| `vocabulary` | object | Concrete nouns grouped by domain. |
| `signaturePhrases` | string[] | Lines that are load-bearing for the brand. |
| `changelog` | object[] | One entry per revision. Each links to the correction that caused it. |
| `voiceQueue` | object[] | Captured corrections, adopted or pending. The input side of the loop. |

## `guidelines[]`

```json
{
  "rule": "Talk like a regular, not a sommelier.",
  "derivedFrom": "Unpretentious",
  "explanation": "Tasting notes help people choose. Tasting poetry makes them feel dumb.",
  "derivedFromCorrection": "vn_0002"
}
```

`derivedFromCorrection` is the join to `voiceQueue`. It answers "why does this rule exist?" with a real correction, not a committee.

## The loop

This is the part that matters. Most brand voice is written once and rots. This schema treats voice as a living artifact:

1. A draft comes back wrong. Someone (a human or an agent) files a note in `voiceQueue`.
2. During a periodic sweep, adopted notes become or amend `guidelines`.
3. `revision` bumps and a `changelog` entry records the change, linked back to the note.
4. The next generation reads the updated spec. The mistake does not repeat.

`data/changelog.jsonl` joins each revision to the correction that produced it, so you can read the voice's whole evolution as cause and effect.

## Using it with a model

- Put `guidelines` + `bannedTerms` + a few `preference_pairs` in the system prompt to steer generation.
- Pick the right `register` for the surface you're writing.
- Use `promptKit` entries as parameterized templates for recurring jobs.
- Score output against `bannedTerms` and the rules as a cheap voice-lint.