Skip to content

Master prompt best practices

The master prompt is your assistant’s operating instructions — tone, scope, escalation rules, and guardrails. It is sent to the Response Synthesizer on every turn, alongside retrieved chunks and tool results.

SectionExample guidance
Role”You are the support assistant for Acme Corp.”
Scope”Answer only from the knowledge base and connected tools.”
Tone”Friendly, concise, professional — no jargon.”
Citations”Cite document titles when answering from the KB.”
Tools”Use lookup_order when the visitor provides an order ID.”
Escalation”If you cannot help, offer to connect them with support@acme.com.”
Rich content”Do not invent prices or product images — only show data from tools.”

You are a helpful assistant for our business. Answer from the knowledge base and connected tools only. Cite sources by name. If information is missing or a tool fails, say so clearly and offer to connect the visitor with our team. Never make up prices, inventory, or offers.

Do

  • Keep prompts under ~500 words — long prompts dilute retrieval context
  • Mirror language your visitors use (“order” vs “purchase”)
  • Define what “I don’t know” looks like for your brand
  • Test edge cases in the test panel

Don’t

  • Paste entire policy documents — put those in a knowledge base
  • Instruct the model to reveal credentials or bypass tools
  • Ask for structured product cards without a connected tool — see Trust policy

Your plan controls available models. Advanced models (Standard+) handle longer prompts and complex tool orchestration better. See Model selection.

Smart query understanding (Starter+) runs before the synthesizer — the master prompt should complement routing, not duplicate it. Avoid lines like “always search the knowledge base first”; let the Query Strategist decide.

  1. Review real conversations in the dashboard.
  2. Add clarifications for repeated failure modes.
  3. Re-test before saving to production deployments.