Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.insitechat.ai/llms.txt

Use this file to discover all available pages before exploring further.

​
Definition

A system prompt is the standing instruction sent to a large language model before every conversation, defining the chatbot’s persona, tone, knowledge boundaries, and behavior. It is the single most important configuration in any LLM-powered chatbot — it determines how the bot introduces itself, what it refuses to do, how it handles questions it can’t answer, and what voice it speaks in. In InsiteChat, the system prompt is set per chatbot in the Instructions tab.

​
What a system prompt controls

AspectExample instruction
Identity”You are Sage, the AI assistant for Acme Dental.”
Tone”Friendly and professional. Use plain language; avoid jargon.”
Scope”Only answer questions about Acme Dental services and policies. Do not advise on legal or medical matters.”
Refusal style”If you don’t know the answer, say so honestly and offer to connect the visitor with a human.”
Format”Keep answers under 3 sentences unless asked for detail. Use bullet lists for steps.”
Language”Respond in the language the visitor uses.”
Brand voice”Reference our ‘patients-first’ philosophy when relevant.”

​
How a system prompt is used at query time

When a visitor sends a message, InsiteChat composes a prompt like this and sends it to the LLM: 
[SYSTEM PROMPT — your standing instructions]
[RETRIEVED CONTEXT — top-K chunks from your content]
[CONVERSATION HISTORY — last N turns]
[VISITOR MESSAGE — the current question]
The system prompt always comes first. It governs every response — even ones that fall back to the LLM’s own knowledge when retrieval fails.

​
Writing an effective system prompt

InsiteChat ships with 6 pre-built templates to save you starting from a blank page:
  • Customer Support — friendly, helpful, escalates to humans gracefully
  • Sales — enthusiastic, persuasive, captures leads at the right moment
  • Technical — precise, code-aware, links to documentation
  • Education — patient, explanatory, uses examples and analogies
  • Enterprise — formal, compliance-aware, careful with claims
  • FAQ — direct, factual, minimal embellishment
Each template is a starting point — edit it to match your brand. Most teams customize 3-5 sentences and ship.

​
Guidelines

  1. Open with identity: “You are [Name], the AI assistant for [Company].”
  2. Set tone in one line: “Tone: warm, professional, no jargon.”
  3. State scope explicitly: “Answer questions about [X, Y, Z]. Decline questions about [A, B, C].”
  4. Handle the unknown: “If you don’t know, say ‘I’m not sure — let me connect you with someone who can help’ and trigger a lead form.”
  5. Stay under 300 words — the LLM weighs every instruction, so brevity helps.

​
What system prompts cannot do

  • They cannot teach the model facts about your business — that’s what RAG is for.
  • They cannot override pricing or product info encoded in retrieved content.
  • They cannot fully suppress hallucinations on questions where retrieval returns nothing — for those, you need custom Q&A pairs with precise answers.

​
InsiteChat’s persona system

In addition to the system prompt, InsiteChat lets you pick a persona (Friendly, Professional, Casual, or custom) that adds tone-shaping instructions to every response. Personas are a fast way to enforce voice without rewriting the full system prompt. See Widget customization for the full list.

​
Learn more