Can I send system role messages mid-conversation?

can i send a "system" role messages in the middle of a conversation using the ai sdk? does that make sense?

Now I have a complete picture. Let me summarize the findings for you:

##Summary: Can you send "system" role messages in the middle of a conversation?

Short answer: The AI SDK allows it at the type level, but it's not recommended and may cause issues with some providers.

###Here's the detailed breakdown:

####1. Type-level: System messages ARE allowed in the messages array

Looking at packages/ai/src/prompt/message.ts and packages/provider-utils/src/types/model-message.ts, the ModelMessage type is defined as:

export type ModelMessage =
  | SystemModelMessage
  | UserModelMessage
  | AssistantModelMessage
  | ToolModelMessage;

So you can include a message with role: 'system' anywhere in your messages array.

####2. However, there's a strong warning in the source code

From packages/provider-utils/src/types/system-model-message.ts:

/**
 * A system message. It can contain system information.
 *
 * Note: using the "system" part of the prompt is strongly preferred
 * to increase the resilience against prompt injection attacks,
 * and because not all providers support several system messages.
 */

####3. Different providers handle this differently

• OpenAI: Simply passes through system messages wherever they appear (packages/openai/src/chat/convert-to-openai-chat-messages.ts) - the OpenAI API does allow system messages at various positions.

• Anthropic: Throws an error if you have multiple system messages separated by user/assistant messages (packages/anthropic/src/convert-to-anthropic-messages-prompt.ts):

  if (system != null) {
    throw new UnsupportedFunctionalityError({
      functionality:
        'Multiple system messages that are separated by user/assistant messages',
    });
  }
  

• Amazon Bedrock: Similar restrictions - combines leading system messages but throws for separated ones (packages/amazon-bedrock/src/convert-to-bedrock-chat-messages.ts)

• Google (Gemma): Combines system messages into the first user message for compatibility

####4. The recommended approach: Use the system parameter

The AI SDK provides a separate system parameter in generateText(), streamText(), etc.:

const result = await generateText({
  model: openai('gpt-4'),
  system: 'You are a helpful assistant.',  // ← Use this instead
  messages: [
    { role: 'user', content: 'Hello!' },
    { role: 'assistant', content: 'Hi there!' },
    { role: 'user', content: 'How are you?' },
  ],
});

This ensures the system message is always at the start and is handled correctly by all providers.

###Does it make sense semantically?

Some use cases where mid-conversation system messages might make sense:

• Dynamically injecting context or instructions as the conversation evolves
• Simulating "internal thoughts" or changing agent behavior mid-conversation

However, not all LLM APIs support this pattern, which is why the AI SDK recommends using the dedicated system parameter for maximum compatibility.