Brand Voice

This skill complements the brand-voice plugin. The plugin runs as post_receive middleware after the assistant finishes a turn: it inspects the final text against banned phrases, banned patterns, required phrases, and an optional LLM classifier, then either allows, rewrites, or blocks the response.

The plugin is a safety net. This skill is the up-front discipline that keeps the safety net from firing.

When to use

Activate this guidance whenever the response will reach an external audience: customers, partners, stakeholders, public comments, marketing copy, sales follow-ups, support replies. It is also a sensible default for internal messages on user-facing channels (Slack to a customer team, email to a vendor).

Working rules

1. Read the configured voice before drafting. Run /brand-voice (the plugin command) to see the active mode, banned phrases, banned patterns, and required phrases. Treat that output as authoritative. If no voice is configured, write in clear neutral business prose and stop second-guessing tone.
2. Lead with substance, not voice. Brand voice is the polish on top of a correct, concrete answer. Never trade accuracy for tone.
3. Avoid the banned set. Do not use banned phrases or anything that matches the banned regex patterns. If a banned phrase is the most natural word, find a synonym or restructure the sentence — do not simply alias the banned word with a similar one.
4. Honor required phrases when contextually appropriate. If the config lists required phrases (e.g. a tagline, disclaimer, or salutation) and the response is a customer-facing message, include them. Do not insert them into purely internal or technical replies where they would feel forced.
5. Mirror the caller's register. If the user is informal, do not reply in stiff legalese; if they are formal, do not reply with slang. Brand voice describes the ceiling and floor, not the exact register.
6. Preserve facts and citations across rewrites. When you self-correct a draft for tone, keep every concrete claim, link, code block, and number unchanged.

Failure modes the guard will catch

• Off-brand jokes, sarcasm, or hedges in customer-facing replies.
• Banned-phrase appearances ("guaranteed", "obviously", competitor names, …) — depends on configuration.
• Missing required disclaimers in regulated contexts.
• Sycophantic openings ("Great question!") or apology spirals when the voice is explicitly direct.

If you notice yourself drifting into any of these, stop and rewrite the draft before sending. The guard's rewrite is a fallback; an in-thread rewrite preserves intent better than a downstream model rewrite.

Plugin interaction

• Mode flag: the guard only logs. Do not relax — drift will still show up in audit reports.
• Mode rewrite: the guard will call the configured aux model to rewrite off-brand output. This is the default, and it is not free — every off-brand response costs an extra model call and adds latency.
• Mode block: the guard replaces the message with a fixed blockMessage. Treat this as a hard error: an off-brand draft means the user gets nothing.

The cheapest, fastest, and highest-fidelity outcome is always: write on-brand on the first pass.