Guidelines & system prompts
Guidelines is the system prompt — the instructions the AI sees on every turn. This is where you tell the bot who it is, how it should sound, and what it must not do. After knowledge sources, this is the highest-leverage knob in the product.
In this guide:
- Open Guidelines
- The structure of a good system prompt
- Examples for common bot types
- Versioning and rollback
- What to test after changing
Step 1: Open Guidelines
On the chatbot detail page, click Guideline in the left rail.
Screenshot: The Guidelines editor with the active system prompt.
Step 2: Understand what’s already there
The editor shows the current system prompt. New chatbots get a sensible default based on the Type you picked in General settings — a Support type ships with a support-flavored prompt; a Sales type ships with a sales one.
Read it before editing. The default is your baseline; only override what’s wrong for your use case.
Step 3: Structure your prompt
A reliable system prompt has four parts:
[ROLE] — Who is the bot?
[TONE] — How does it talk?
[SCOPE] — What can / can't it answer?
[RULES] — Hard prohibitions and required behaviors.Example:
You are Pip, the support assistant for Acme Inc.
Tone: friendly, concise, no exclamation points unless apologizing.
Scope: answer questions about Acme products, billing, and orders.
Refuse politely if asked about competitors, legal advice, or anything
unrelated to Acme.
Rules:
- Never invent product features. If unsure, say "Let me check with a human."
- Never share order details without confirming the customer's email matches.
- Always offer to escalate to a live agent if the user types "human", "agent",
or "talk to someone".Step 4: Examples for common bot types
Customer support
You are <Bot Name>, the support assistant for <Company>. Be friendly,
concise, and accurate. Use only the knowledge sources available; never
guess. If you don't know, offer to escalate. Confirm the user's identity
(email match) before sharing any order or account specifics.Internal IT help desk
You are <Bot Name>, the IT help bot for <Company> employees. Use a clear,
professional tone. For password resets, billing questions, or urgent
production issues, immediately escalate to the on-call engineer.E-commerce sales assistant
You are <Bot Name>, the shopping assistant for <Store>. Help shoppers find
products, compare options, and complete checkout. Always cite the product
URL when recommending. Never invent prices or stock — defer to live store
data via actions.Step 5: Save and test
Click Save. The change applies to new messages only — in-flight conversations keep their previous prompt.
Open the Playground and re-run your test sweep. Watch for:
- Tone shifts (the bot now sounds less like the old version).
- Refusal patterns (does it correctly decline out-of-scope questions?).
- Action triggers (do “talk to a human” phrases escalate as expected?).
Versioning and rollback
Hilal Chatbot keeps a history of saved Guidelines. Click Version history in the editor to view previous versions and Restore to roll back. Restore creates a new version (it doesn’t truncate forward history) so you can always re-roll-forward.
Common mistakes
- Telling the bot to “always be helpful, accurate, and friendly.” The model already wants to do that. Generic instructions add tokens but no signal.
- Putting product knowledge in the system prompt. Use Knowledge sources — they’re indexed, citable, and don’t blow your context budget.
- Forgetting refusal rules. “Don’t talk about competitors” prevents awkward conversations. Be explicit.