Default assistants sound like default assistants: competent, generic, slightly eager. The temptation is to overcorrect with a mascot — a name, a joke density, a catchphrase on every refusal. That usually makes the product feel less trustworthy, especially at the moment it has to say no.
Persona work that helps is quieter. Greeting, register, and boundaries. The how-to is in tune your bot persona; this post is what to aim for so you do not confuse brand voice with entertainment.
The greeting is a scope contract
“Ask me anything” invites questions outside your sources and manufactures refusals. “Ask about shipping, returns, or sizing” produces answerable traffic. The greeting is not marketing copy — it is the first filter on whether the conversation can succeed.
Match the site you already shipped
Read your homepage and help centre aloud. If they are plain and specific, the assistant should be too. If they are warm and short, keep replies warm and short. An assistant chattier or more formal than everything around it reads as bolted on, which undercuts the credibility citations are supposed to build.
Rules beat vibes
Write the categories it must not attempt: medical or legal advice, account-specific balances, custom quotes, delivery promises it cannot see. For each, say what to do instead — escalate, point at a form, or decline. Vague instructions (“be careful”, “use judgment”) fail under paraphrase pressure; named categories hold.
Knowledge still comes from indexed sources. Persona does not teach facts. It decides manners and declines. If a fact is wrong, fix the page or the crawl — do not paper over it with a friendlier system prompt.
Where mascots go wrong
- Jokes on top of a refusal feel like mockery when the customer is stuck.
- Forced first-name chumminess clashes with regulated or clinical brands.
- Long “brand stories” in the system prompt dilute the rules that matter.
- Inconsistent register across languages makes multilingual setups feel broken.
If your brand truly has a character — a cafe, a kids’ product, a playful DTC voice — use it lightly and test refusals with real customers. The character should survive an “I don’t know” without needing a rimshot.
Test boundaries, not banter
- Ask three in-scope questions; check citations and tone against the site.
- Ask three out-of-scope questions; confirm plain refusals and next steps.
- Ask for a human; confirm immediate handover language matches the team.
- If you support multiple languages, repeat the refusal test in each.
Connect persona work to helpful refusals and escalation. Tone without a path to a person is decoration. Boundaries without tone still work; tone without boundaries invents.
Ship the boring version first
A clear, on-brand, slightly dry assistant that refuses well will outperform a memorable mascot that improvises. You can always add warmth later. You cannot easily claw back trust after a funny wrong answer about refunds.
