Voice to system prompt

"Voice" is not a vibe string. It is structured language instruction with exactly one consumer: buildBrandSystemPrompt(), whose output becomes the agent's instructions (and, in the mock, the rules draftReply() follows).

The model

voice: {
  tone: 'Friendly, plain-spoken local pro' // one line, human-readable
  formality: 'casual' | 'professional' | 'formal'
  dos: ['Lead with the fix, not the apology']
  donts: ['Never quote a price in a reply']
  examples: [{ situation: 'question', reply: '…' }] // 2–3 gold replies
}

• tone is the headline a human reads on the Brand tab.
• formality is the machine-readable dial: greeting style, contraction use, sign-off length.
• dos / donts are hard rules — short imperative lines, each independently enforceable.
• examples are few-shot gold replies keyed by post type (question, complaint, praise). The agent mimics these before it improvises.

The compiler

buildBrandSystemPrompt(brand) is deterministic and versioned (PROMPT_VERSION = 1):

1. Identity line — who the agent speaks as.
2. Tone + formality rules.
3. Offerings list with one-line details.
4. Dos / don'ts, verbatim.
5. Gold examples, in full.

Every draft records the prompt version that produced it, so a voice edit never silently rewrites history — drafts stamped v1 stay v1.

Preview

The Brand tab previews the compiled prompt per channel through the Test it simulator: type an inbound message, get the exact raw reply. What you see there is what the agent will send. If the preview reads wrong, the channel snippets are wrong — fix the snippets, not the drafts.

Channels and memory

Global voice is only half the story. Each channel (facebook, x, reddit) carries its own ChannelProfile — casual vs standard style, raw texting snippets, and (on Facebook) the triage flow the agent pushes for first. The memory bank (memory.rules) compiles into the prompt as working facts the agent must never contradict.