Personality + custom instructions

The personality is the core paragraph that tells the AI who it is, how it should sound, and what it cares about. The custom instructions are additional rules appended after the personality.

Personality

Length: one paragraph (100–300 words).

Structure:

Role. “You are a…”
Tone. Friendly, formal, enthusiastic, measured.
Priorities. What to do first when a visitor speaks. Answer quickly? Ask clarifying questions? Recommend?
Don’ts. Things to avoid saying or doing.
Escalation path. When to say “let me get a human” or share a phone number.

A good example

You are Maya, the AI host at Ember & Oak, a farm-to-table restaurant in
downtown Los Angeles. You speak with warmth and genuine enthusiasm — you
love this food and it shows. Answer concisely and confidently when you know
something; freely admit when you don't. Help guests find a dish they'll
love by asking one or two clarifying questions about preferences and
dietary restrictions.

Always mention if a dish has common allergens (nuts, dairy, shellfish,
gluten) when a guest asks about it. Do not invent specials or promotions
that aren't on today's menu. If a guest wants to make a reservation, point
them to the "Reservations" button on the page — you can't book for them.
If a guest sounds upset about their experience, apologize briefly and
offer to connect them to the manager at (555) 123-4567.

Why this works

Named character — “Maya” creates a sense of continuity across visits.
Specific tone — “warmth and genuine enthusiasm” is stronger than “friendly.”
Concrete don’ts — “don’t invent specials” prevents hallucinations on a common attack surface.
Fallback to humans — the manager’s phone number is in the prompt so the AI doesn’t have to invent one.

What not to do

Vague directives. “Be helpful” — meaningless.
Contradictions. “Always be concise but explain thoroughly” confuses the model.
Too many rules. The model stops following rule 15 after rule 5 changes its mind. Keep it tight.
Marketing copy. The personality is for the AI, not for visitors. “Best in LA” tells the AI to brag — don’t.

Custom instructions

Appended after the personality, more rule-like. Good for:

Compliance text. “Never discuss individual patient records.” (Healthcare)
Pricing guardrails. “If asked about discounts, always check the current site promotions before committing to a number.”
Escalation rules. “If a visitor asks about legal advice, say ‘I can share general information but a licensed attorney needs to advise on specifics.’”
Language/locale rules. “Use metric units by default.”

Compliance rules:
- Never discuss individual patient records, insurance claims, or appointment times.
- If asked about medical symptoms, give general information only and recommend
  consulting a doctor. For urgent symptoms (chest pain, difficulty breathing,
  severe injury), immediately instruct the visitor to call 911.
- Do not quote specific prices — refer to the "Pricing" page.

Variables you can use

Token	Replaced with
`{business_name}`	Your site’s `business_name`
`{business_description}`	Your site’s `business_description`
`{current_time}`	Resolved at session start (in your timezone)
`{today}`	`Friday, April 17, 2026` format
`{is_open}`	`true` or `false` (see business hours)

Example:

Today is {today} and it's {current_time}. {is_open ? 'We are open.' : 'We are closed.'}

Inheriting from a template

When you select an industry template, the personality field is pre-filled. Use that as your starting point and edit.

Changing mid-session

The personality is loaded at session start. Changing it in the dashboard won’t retroactively affect live sessions (they keep their original prompt). New sessions use the updated one.

Length impact on cost

Longer personalities use more tokens per session. Rough math: every 100 tokens of personality costs ~$0.0001 per session. A 400-token personality on a Pro-plan site (2000 min/mo ≈ 4000 sessions) adds ~$1.60/month. Negligible.

Testing

Dashboard → Voice → Preview prompt → see the full resolved prompt → Test voice → try it out.