When you work with an LLM, it is really important to set the tone and capabilities of these amazing tools.

That’s because LLM are incredibly powerful, but they start as generalists and you let them figure out what the focus of the conversation will be by understanding the conversation.

This is not a practical use of our time, nor the tokens!

You have many way to try to setup a system prompt out in the wild, you can try to pass it as a first message of the conversation, but thanks to the AI SDK you have a very specific approach. All you have to do is to pass your system prompt, the string where you describe what you want and how from the LLM, right inside a the streamText configuration object.

const SYSTEM_PROMPT = `
You are a very skilled developer and you know all the things about TypeScript.

ALWAYS convert the code you read into TypeScript.

Do your best to generate types that are understandable and do not duplicate code. If a type has the same shape of another type, or a part of, do your best to derivate it and use the original as the source of truth.
`

const streamTextResult = streamText({
  model: google('gemini-2.5-flash'),
  messages: modelMessages,
  system: SYSTEM_PROMPT
})

Honestly, there’s nothing more for this lesson.

Start by experimenting with your own needs and see how much you can force the LLM to behave as you whish.