Skip to main content

​
What instructions do

Instructions are the system prompt for your agent. They’re the highest-leverage thing you can edit — every response is shaped by them. Halo’s engine combines your instructions with:
  • The current user’s profile, company, and context
  • Knowledge base search results
  • Conversation history
  • Tool definitions for the agent’s enabled skills and actions
Then sends it all to the model. Get the instructions right and everything downstream improves.

​
Where to edit

AI Agents > [Your Agent] > Configure, in the Instructions field. Halo also generates a portion of the system prompt automatically — your data sources, brand voice, and structured fields like “Primary Goal” and “Boundaries” — so you don’t need to repeat those.

​
What good instructions look like

Aim for three things: role, rules, and style. 
You are the Halo AI support agent for Acme. You help customers troubleshoot
their integrations, answer product questions, and create support tickets when
human help is needed.

Rules:
- Always check the user's connected integrations before suggesting reconnection.
- Never quote pricing — direct billing questions to [email protected].
- If the user mentions a bug, file a Linear ticket using the create_bug action.

Style:
- Be concise. Three sentences or fewer for simple questions.
- Use the user's first name when greeting.
- If you're unsure, say "I'm not sure — let me get a teammate" instead of guessing.

​
What to avoid

  • Don’t repeat what’s already in your knowledge base — the AI searches knowledge automatically. Adding “We have a feature called X” in instructions duplicates the docs and goes stale.
  • Don’t list every product feature — instructions get truncated; knowledge search handles facts.
  • Don’t include user-specific details — those come from traits/context dynamically.
  • Don’t try to encode the entire support process — use channels, escalation rules, and skills for that. Instructions are for how the AI talks, not what tools it has.

​
Structured fields (Data Sources tab)

The Data Sources tab adds three structured prompts on top of your instructions:
FieldPurpose
Primary GoalOne-sentence mission. The AI re-orients to this whenever a conversation drifts.
Ideal CustomerWho you serve. Helps qualifying questions and tone.
BoundariesThings the AI must never do (e.g. “never give legal advice”).
These are appended to your instructions and applied with stronger weighting in the system prompt. Use them for invariants you want the AI to follow consistently.

​
Onboarding mode

Each agent has an Onboarding config (under Configure) that lets you set:
  • Onboarding steps — an ordered list of things the AI should walk a new user through (“Connect your Stripe account”, “Create your first form”).
  • Exit conditions — when to stop being in onboarding mode (e.g. “user has connected at least one integration”).
  • Onboarding-specific instructions — added on top of the regular system prompt while the user is in onboarding.
When a user is in onboarding mode, the AI prioritizes guiding them through the steps over answering general questions.

​
Human mode

Human mode controls how “human-like” the agent feels:
  • Typing speed and response delays — make replies feel typed instead of instant.
  • Persona name — shows in the chat (“Matt is typing…”) if you want the AI to identify as a teammate rather than a bot.
  • Fillers, emojis, small talk — toggles for casual conversational moves.
Use this when you want a warmer, more conversational feel. Skip it for B2B SaaS where speed matters more.

​
Iterating

The fastest way to improve instructions is to:
  1. Read transcripts in Inbox > Conversations for ones where the AI under-performed.
  2. Spot the pattern — was it tone, missed knowledge, hallucinated information, didn’t escalate?
  3. Patch the gap:
    • Tone or style → update instructions
    • Missed knowledge → add a knowledge source
    • Wrong tool used → re-check skills/actions
    • Should have escalated → tighten escalation rules
The Train tab lets you practice in a sandbox without affecting real customers.

​
Where to go next

Skills

Turn on capabilities the agent can use.

Live Train

Practice and iterate on instructions without touching production.