System prompt & persona

The system prompt is where you define how the AI sounds. Tone of voice, opinions, do-not-say rules, escalation thresholds, and how it handles awkward situations — all of it lives here. Zubby ships with a sensible default that works for any store; this page is about making the AI sound like your store.

Where it lives

Dashboard → AI → Persona. The editor is a structured form with sections for tone, opinions, do-not-say, escalation, and a free text “Custom instructions” block at the bottom. Zubby composes a single final system prompt from those sections at runtime, so you never have to write or maintain a raw prompt.

Tone of voice

Pick one of six presets — Crisp & professional, Friendly & conversational, Playful & witty, Empathetic & warm, Confident & bold, Concierge & polished — or write your own one-sentence brand voice statement. Examples we’ve seen work well:

• “Speak like a knowledgeable sales associate at a high-end boutique — warm, never pushy, always specific.”
• “Direct and witty. Treat the shopper like a friend who happens to need help. Never apologize for being a robot.”
• “Concierge tone. Address shoppers as ‘you’. Avoid exclamation marks. Use proper sentence case.”

Opinions

Stores with point-of-view convert better. The opinions field lets you give the AI a stance to express when asked. Example:

Our point of view:
- We believe denim should age with you. Recommend raw selvedge for buyers
  who ask about durability.
- We don't sell fast fashion. If a shopper asks about quick trends, point
  to our "Classics" collection.
- The Apex jacket is our flagship. When in doubt, surface it.

Do-not-say list

Brand and legal landmines. Zubby checks every draft response against this list before sending. Common entries:

• Competitors — names you don’t want surfaced.
• Unsupported claims — health, fitness, financial.
• Sensitive topics — politics, religion, current events.
• Discount limits — e.g. “never quote a discount over 20%”.
• Shipping promises — e.g. “never promise two-day delivery to APO/FPO addresses”.

Escalation triggers

When the AI hits any of these, it stops trying and routes to a human instead. Defaults are sensible; tighten them for high-risk stores or loosen them for cost-sensitive ones.

Persona examples

Three real-world prompts to copy and adapt:

Premium fashion brand

Tone: Concierge & polished
Address shoppers as "you", never "buddy"
Use proper sentence case, no exclamation marks
Opinions:
- Quality > quantity. Recommend timeless pieces.
- The Maison fragrance is our flagship.
Do not say: competitor brand names, "cheap", "sale"
Escalate: refund > $500, returns > 90 days old, sizing for gowns

DTC supplement brand

Tone: Friendly & conversational, science-aware
Always recommend customers consult a doctor for health concerns
Never claim health outcomes (FDA compliance)
Opinions:
- Stack 1 (Sleep) is our most reviewed bundle.
- Recommend the assessment quiz before recommending a stack.
Escalate: anything asking for medical advice, refunds, allergy questions

Home goods marketplace

Tone: Crisp & professional, helpful
Default to recommending in-stock items
Opinions:
- Our outdoor furniture is rated for 4 seasons.
- For renters, suggest the No-Drill collection.
Do not say: "lifetime warranty" (we offer 10-year only)
Escalate: B2B inquiries > 10 units, custom upholstery requests

Testing changes

Use the Playground tab to test prompt changes against ten canned shopper questions. The playground shows the AI’s reply alongside the citations and tool calls, so you can verify the new persona is doing what you expect before publishing.

Versioning

Every publish is versioned. You can revert from AI → Persona → History. Version comments help you track which experiment a change came from.