Skip to main content
Build is the area of the Chatbase dashboard where you configure how your AI agent works: what it says, what it knows, and what it can do. Build > Instructions holds your AI agent’s written instructions, its AI model, and its temperature, which together decide how it answers across your channels, from the Chat Bubble to WhatsApp and email. It’s for whoever owns the AI agent’s behavior, such as a support lead, marketer, or developer. Change your instructions when the AI agent’s tone, scope, or rules don’t match what you want, for example when it answers off-topic questions or doesn’t offer a handoff. Change the model when answers are accurate but too slow, too shallow, or too costly in message credits. If answers are wrong because information is missing, fix your Data Sources instead. This page covers the Instructions section: the instructions themselves, the model, the temperature, and the Compare view for testing changes. The other Build sections have their own pages:
Two different things are called Guardrails. The Guardrails instructions field on this page holds behavioral rules you write for the AI agent. The separate Guardrails page covers abuse protection: rate limits, spam detection, and country blocking.

​
Instructions

Instructions define your AI agent’s behavior, role, and response guidelines. Use them to provide context, set expectations, and control how your AI agent responds to users. Go to Build > Instructions. The instructions editor fills the left of the page, and the Model configuration panel on the right holds the Model dropdown and the Temperature slider. Use Compare in the header to test changes side by side, and Save changes to apply them.
Saving puts your changes in a draft. Your customers see them once you click Publish in the top-right corner. A few settings go live as soon as you save. See Drafts and versions.

​
General instructions

The General section contains your AI agent’s primary instructions. Use this section to define:
  • Your business or product context
  • The AI agent’s role and responsibilities
  • Response style and tone
  • Goals and expected behavior
  • Any additional guidance the AI agent should follow
These instructions serve as the foundation for your AI agent’s responses. Want links your AI agent shares to keep a visitor’s campaign UTM parameters, like utm_source? See UTM Tracking for the instructions to add.

​
Guardrails instructions

Use Guardrails to define rules your AI agent should always follow. Examples include:
  • Restricting topics the AI agent can discuss
  • Preventing the disclosure of sensitive information
  • Defining safety or compliance requirements
  • Controlling fallback behavior
  • Limiting the AI agent’s role or capabilities
Guardrails are intended for non-negotiable rules that should apply to every conversation.
Start with a simple, clear description of your AI agent’s purpose, then gradually add specific instructions based on user feedback and common queries. For a full worked example you can copy, see Best Practices.

​
Instruction examples by use case

Good instructions name the business, the AI agent’s job, what it should and shouldn’t do, and what to say when it doesn’t know. Replace the sample company details below with your own.
General
Guardrails
Instructions only describe what the AI agent should do. To let it book a demo, look up an order, or start a return, add the matching Actions, such as Calendly, Shopify Actions, or Escalations. For flows with several ordered steps, such as returns, use a Procedure.

​
AI Model

The AI model is the specific machine learning model used to generate responses for your AI agent. Each model has different capabilities, performance characteristics, and message credit cost, so you can choose the one that best fits your needs.
Auto adapts to every conversation, giving instant answers for everyday questions and frontier-grade intelligence for demanding ones. Multilingual and image understanding are built in, and it handles tools end to end.
  • Auto: 1 message credit
  • GPT-6.1 Sol: 2 message credits
  • GPT-6.1 Terra: 2 message credits
  • GPT-6 Luna: 1 message credit
  • GPT-5.6 Terra: 2 message credits
  • GPT-5.5: 4 message credits
  • GPT-5.2: 2 message credits
  • GPT-OSS-120B: 1 message credit
  • GPT-OSS-20B: 1 message credit
  • Claude 5.5 Opus: 5 message credits
  • Claude 5.5 Sonnet: 2 message credits
  • Claude 5.5 Haiku: 1 message credit
  • Gemini 3.8 Flash: 2 message credits
  • Gemini 3.6 Flash: 2 message credits
  • Gemini 3.5 Flash Lite: 1 message credit
  • Gemini 3.1 Pro: 2 message credits
  • Gemini 3.1 Flash Lite: 1 message credit
  • Gemini 3 Flash: 1 message credit
  • Gemini 2.5 Pro: 2 message credits
  • DeepSeek-V4.1-Flash: 1 message credit
  • DeepSeek-V4-Flash: 1 message credit
  • DeepSeek-V3.1: 1 message credit
  • DeepSeek-R1: 1 message credit
  • Grok 4.7: 2 message credits
  • Mistral Medium 3.5: 2 message credits
  • Mistral Small 4: 1 message credit
When you choose a Mistral model, the main inference call that generates your AI agent’s response runs on servers in the EU. Other steps involved in producing and delivering the reply may still be processed outside the EU, so selecting a Mistral model does not by itself keep all conversation data in the EU.
  • Kimi K2.6: 1 message credit
  • GLM 5.3 Flash: 1 message credit
  • GLM 5.2: 2 message credits
For help choosing between them, see the models comparison page.

​
Temperature

Temperature controls how creative and varied your AI agent’s responses are. The setting ranges from 0 to 1 and directly affects response predictability:
  • Lower temperature (close to 0): Produces focused, consistent responses by selecting the most probable outputs
  • Higher temperature (closer to 1): Generates more creative and varied responses with less predictability
Drag the slider between Reserved and Creative to set the value.
The default temperature is 0. Adjust it to experiment with different response styles and find what works best for your use case.

​
How temperature changes answers

Temperature changes how an answer is worded, not what the AI agent knows. Suppose a customer asks “What are your opening hours?” and your Data Sources say you’re open 9 AM to 6 PM, Monday to Friday:
  • At 0, the AI agent gives much the same answer each time: “We’re open Monday to Friday, 9 AM to 6 PM.”
  • Near 1, the wording changes from one conversation to the next: “Happy to help! You can visit us any weekday between 9 in the morning and 6 in the evening.”
Keep temperature low for support, billing, and policy questions, where customers should get the same answer every time. Raise it if replies feel repetitive in conversations where variety helps, such as product recommendations. To see the difference with your own content, open Compare and give two chats the same model at different temperatures.

​
Compare area

Click Compare in the Build > Instructions header to open the compare view.
Instructions Compare view
The Compare area lets you put several versions of your AI agent side by side and give each its own settings, which makes it much easier to work out which settings suit you. The same message is sent to every chat, so you can test the AI agent’s response to identical input under different settings.

​
Per-agent controls

Each chat box in the compare view has its own controls, from left to right:
Instructions Compare view actions

​
View controls

You can also:
  • Clear all chats
  • Reset the AI agent’s settings to the main settings
  • Add a new AI agent to test with
  • Click the filter icon to change which AI models you’re testing, and select a new model from the dropdown
Send the same question to two or three models at once and compare the answers before you commit to one in Model configuration.

​
Troubleshooting

Possible cause: you’ve run out of message credits.Solution: purchase add-on message credits. See Add-ons for details.

​
Frequently asked questions

Instructions apply to every conversation: the AI agent’s role, tone, scope, and rules. A Procedure applies only when a conversation matches its trigger, and gives the AI agent an ordered list of steps to follow, including the Actions to call. Use instructions for how your AI agent behaves in general, and Procedures for situations like refunds or order issues where you don’t want it improvising.
Instructions can be up to 30,000 characters, whether you write them in the dashboard or set them through the Chatbase API. Shorter instructions are easier to maintain: put facts like prices and policies in your Data Sources, and keep instructions for behavior.
Yes. In the Playground, the Overview tab of each channel has Sync with global instructions. Leave it on to use the instructions from Build > Instructions, or turn it off to write instructions for that channel only.
Saving puts your changes in a draft. The Playground uses the draft right away, but your live channels keep the published version until you click Publish. See Drafts and versions.
If you’re not sure, start with Auto, which uses 1 message credit per reply and adapts to each conversation. To compare specific models, send the same questions to several of them in Compare, and see Models comparison for each model’s strengths.
Put context and everyday behavior in General: your business, the AI agent’s role, tone, and goals. Put non-negotiable rules in Guardrails, such as topics to avoid, information never to share, and compliance requirements. The Guardrails instructions field is different from the Guardrails page, which covers rate limits, spam detection, and country blocking.

​
Next steps

Data Sources

Add and manage the content your AI agent is trained on.

Models comparison

See how the available models differ before you commit to one.

Best Practices

Proven ways to improve your AI agent’s answers.

Playground

Preview the effect of your changes the way a visitor will see them.