What they are
Two lists. You write one rule per line, in plain English.
- Do's — behaviors Salva should always exhibit (tone, asks, recommendations)
- Don'ts — behaviors Salva should never exhibit (topics, words, promises)
There's no special syntax. Write it the way you'd brief a new front-desk hire.
Good examples
Do's
- Always confirm the patient's preferred provider before booking.
- Mention that we validate parking when a patient asks about location.
- Ask if it's a new patient — they get a longer first appointment.
Don'ts
- Don't quote insurance coverage amounts. Tell the patient our billing team will verify.
- Don't promise same-day appointments unless the patient is calling about a true emergency.
- Don't mention our pricing for cosmetic services. Always offer a consultation instead.
What this is actually doing
These rules are injected into Salva's system prompt on every call and chat. They sit alongside your FAQs and your AI's greeting, so every response is shaped by them.
Salva uses sensible defaults: polite, professional, defers on clinical questions. Sometimes makes choices that aren't aligned with your practice's preferences.
What "specific" looks like
| Vague (weak) | Specific (strong) |
|---|---|
| Be friendly | Greet returning patients by name when possible |
| Don't be pushy | Don't suggest add-on treatments unless the patient asks |
| Handle insurance carefully | Don't quote coverage amounts; tell the patient we'll verify with their insurer |
The more concrete the rule, the more reliably Salva follows it.
When rules conflict
If a do and a don't could both apply (rare), the don't wins. Salva is designed to fail safe — not saying something is always lower-risk than saying the wrong thing.
What they can't do
Rules can't change Salva's core safety behavior:
- It will always escalate emergencies, regardless of any rule.
- It will not give medical advice, regardless of any rule.
- It will not impersonate a human ("Yes, I'm a person") — it identifies as an AI when asked.
For anything beyond shaping conversational behavior, you're probably looking at AI feature toggles instead.
Published May 2, 2026