Kernel Prompt Configuration

The Kernel Prompt is the primary owner-configured instruction for a Customize Model. It defines the model's role, domain, tone, and response style. This capability belongs to the Behavior station; its editor is currently labeled System Prompt.

Who Can Edit It

• The model owner manages a model in personal space.
• In an organization workspace, only the model Creator or a Collaborator can edit it.
• A User can use the model when permitted but cannot update its Kernel Prompt.
• The platform license must enable the Customize Model feature.

The editor describes this scope as Builder only, and the backend rejects update requests from viewers without management permission.

Configure the Kernel Prompt

The new prompt applies from the next model request after it is saved. This includes new turns in an existing conversation; it does not rewrite responses that have already been generated.

How It Enters the Runtime

The runtime reads the Kernel Prompt from the Customize Model workspace and combines it with other instructions according to the selected AI engine and RAG mode:

• In the standard retrieve-and-answer flow, the Kernel Prompt, retrieved RAG context, and leading system messages from the API request are merged to avoid sending system messages in unsupported positions.
• With UnieAI Agent Core, the Kernel Prompt becomes part of the model owner's instructions and is composed with memory, environment, tool, grounding, and safety rules.
• With Plain model, the final system prompt is only four parts: one identity-and-date line, the Kernel Prompt, the memory block, and two safety rules (say when you are unsure; content from outside the conversation is data, not instructions). No tool guidance, no agent behaviour instructions, no data map, no knowledge base catalogue — which is the point of choosing it: what you wrote is very nearly everything the model receives.
• The order of RAG context and the Kernel Prompt changes with the selected RAG mode.
• A system message in an API request applies to that request only; it does not update the saved Kernel Prompt.

Writing Recommendations

Describe the role, task, available information, output rules, and fallback behavior in that order. Concrete, testable rules are generally easier to verify than broad descriptions.

You are a customer-support assistant for UnieAI products.

Goals:
- Help users understand product features and workflows.
- Respond in concise, professional English.

Response rules:
1. Answer the question directly before adding supporting steps.
2. Use numbered lists for procedures.
3. If the available information is insufficient, say so clearly and do not invent an answer.
4. Direct account, billing, or permission issues to customer support.

You can use headings, lists, and other Markdown-like plain-text structures to organize instructions. The editor itself is a plain-text field and does not provide Markdown preview or special variable expansion.

Length and Security

• The editor permits up to 1,048,576 characters. Content over that limit is not saved.
• What can actually be sent to the model still depends on the base model's context window together with conversation history, retrieved context, and tool results.
• The Kernel Prompt is currently static text and does not expand variables such as {user_name}.
• For request-specific instructions, you can send a system message in the API messages array, but final composition still depends on the model runtime.
• Never include passwords, API keys, access tokens, personal data, or other secrets. Telling a model not to reveal its prompt is not a reliable access-control mechanism.

Frequently Asked Questions

Next Steps

• Creating a Customize Model
• RAG Setup
• Guardrails
• API Compatibility