Shape your Agent guidance
Configure personality, answer length, language, rules, moderation, and follow-up behavior without duplicating facts.
Guidance controls how the Agent answers rather than what the company knows. The current page includes personality, answer length, emoji behavior, optional answer-language enforcement, named answer rules, moderation actions, and follow-up emoji behavior.
This guide explains how to write durable guidance, use the automatic save state safely, and test behavior without turning Guidance into a second Knowledge base.
How this fits into Humind
AI Agent is the configuration space for the customer-facing assistant. Knowledge and Catalog provide facts; Guidance shapes response behavior; Tools add actions; Escalation defines human support; Test surfaces let you review the result before deployment.
Changes can affect many conversations, so test representative buying and support scenarios after each meaningful update. A visually correct chat is not enough: verify the answer, product context, available action, and handoff behavior together.
Before you start
Access: Agent configuration write access is required. Operators and supervisors do not receive this area by default.
- Approve the brand voice and escalation policy with the relevant owners.
- Make factual policies and product data current in Knowledge and Catalog first.
- Prepare paired test prompts that should and should not trigger each new rule.
Step-by-step workflow
Choose the baseline answer behavior
Open AI Agent, then Guidance. Choose the personality that best represents the brand and the answer length that fits the customer experience: concise, standard, or meticulous. Decide whether answer and follow-up emojis are appropriate.
These controls apply broadly. Test several journeys before deciding that a single polished example represents the whole Agent.
Set language behavior deliberately
Enable answer-language enforcement only when the Agent must always use a selected language. The current selector offers English, French, Spanish, German, and Italian. Leave enforcement off when the operating policy is to respond according to shopper context, then test supported languages explicitly.
This setting does not translate Knowledge or catalog data. Content readiness remains a separate responsibility.
Create specific answer rules
Each answer rule has a title, an instruction, and an enabled state. Write a title that names the behavior and an instruction that is direct, testable, and free of unnecessary exceptions. Disable a rule when reviewing its effect instead of creating a contradictory rule.
Keep policies, prices, shipping windows, and other maintainable facts in their source systems. A rule can explain how to handle missing facts or when to escalate, but should not copy the full policy.
Configure moderation
Review offensive, spam, and off-topic behavior. Each criterion can serve or decline according to the available moderation actions. Align these decisions with brand and safety policy, then test clear and borderline examples.
Moderation is not a substitute for support escalation. Decide separately whether a legitimate but sensitive support request should receive human help.
Wait for autosave and run regression tests
The Guidance page saves changes automatically after edits and shows saving or saved status. Wait for saved confirmation before leaving the page or starting a result comparison.
Use fresh Playground conversations and a Batch testing dataset. Test the desired behavior, nearby cases, language, moderation, product recommendations, and support requests to detect unintended broad effects.
Permissions and important caveats
- Autosave means experimental edits can become active quickly; make one controlled change at a time.
- Personality and answer length are global behaviors, while answer rules should remain narrow and explicit.
- Language enforcement does not create translated source content.
- Moderation and human escalation are separate configuration decisions.
Verify the result
Use this checklist before considering the work complete:
- Every enabled rule has a clear owner, purpose, and positive or negative test pair.
- The saved status appears after the final edit.
- Factual answers still come from Knowledge or Catalog rather than copied rule text.
- Brand, language, moderation, product, and support regression scenarios pass.
Troubleshooting
A rule appears to have no effect
Confirm it is enabled and saved, rewrite it as a direct testable instruction, remove conflicting rules, and start a fresh Playground conversation with a clearly matching case.
The Agent follows the rule too broadly
Narrow the condition or scenario described in the instruction. Add a nearby non-matching regression case and avoid generic language that applies to every answer.
The Agent uses the right tone but the wrong fact
Do not keep adjusting Guidance. Find the factual source in Knowledge, Catalog, or an integration, correct it, remove conflicts, and retest the same question.