Learn/Brand Voice Agent
Designing a brand voice agent
A voice agent is the smallest possible brand operating system: one job, one model, one set of inputs. Get this one right and the rest of the system is mostly the same pattern at larger scale.
What a voice agent is
A model with three inputs: voice samples, policy, and the task. It produces drafts that sound like your brand wrote them, refuses requests that violate policy, and asks clarifying questions when the task is ambiguous.
That's it. Everything else is configuration.
The voice samples
Twenty to fifty real pieces of writing you'd happily ship again. Mix formats: a landing page, a launch email, a tweet, a product description, a customer-support reply. Cover the range you'd want the agent to cover.
Don't include adjectives ("we're witty but warm"). Include sentences. The model learns cadence from examples, not labels.
The policy
A short list of things you don't say. Examples that work:
- "Never use 'unlock,' 'leverage,' or 'seamless.'"
- "Don't compare us to competitors by name."
- "Don't make claims about the future tense ('will,' 'going to') without a citation."
- "Default to second person, never first-person plural."
Five to ten items. More than that and the model starts ignoring some of them.
The refusal pattern
Most agents are too eager. They produce off-brand drafts when they should ask for more context. A good voice agent has a refusal pattern: if the task is too vague, it asks one specific question instead of guessing.
Example: ask it for "a tweet about our launch" with no other context, and instead of generating something generic, it should ask "what's the one thing you want a reader to do after seeing this?" before drafting.
Where it breaks
- Stale samples. If your voice has shifted and the samples haven't, the agent sounds like your last brand. Refresh samples when you shift.
- Conflicting policies. "Be conversational" and "be authoritative" can co-exist in human writers; agents struggle. Pick the dominant one and let the human edit the rest.
- Format mismatch. A voice agent trained on landing pages will write a customer-support reply like a landing page. Scope the format or fork the agent.
Where it doesn't break
If you fed it well, the agent holds the voice across hundreds of drafts without drifting. That's the property you can't get from a guidelines PDF: durability. The PDF is read once and forgotten; the agent is consulted every time someone writes.
The leverage
One brand lead can supervise an agent producing hundreds of drafts a week. They review the policy violations, refresh samples quarterly, and otherwise step out of the queue. That ratio is what makes a voice agent the highest-leverage piece of a brand operating system.
Try it in Brandex
Go from concept to a working brand system in minutes.
Frequently asked questions
Can it write headlines, long-form, and customer support — or do I need three agents?⌄
Start with one, scoped to the three formats you ship most. Once it's reliable, fork it. The shared input is voice samples; the per-agent input is format rules.
How often do I refresh the voice samples?⌄
Quarterly is plenty for most brands. More often if you've made a real voice shift; less often if you're stable. The signal to refresh is that the agent's drafts start feeling dated to you.
What model should I use?⌄
Whatever's cheapest that holds your voice. The samples and policies do most of the work; the model is the engine, not the brain. Start cheap; upgrade only if quality plateaus.
Do I need to fine-tune?⌄
Almost never. Retrieval + policy gets you 90% of the way; the remaining 10% is usually not worth fine-tuning cost. Reconsider only at very high volumes.