Assistant prompt is ignored when generating completions

Prompting
, ,
1

I am trying to provide a completion engine with a starting sequence in my prompt, here is the messages structure:

[{
  role: "system",
  content: `You are a software engineer. Given the user input, provide the React app code that solves the task. Example:
<path>app.jsx</app>
<component>
export default function App() {
  return (
    <p>Hello!</p>
  );
}
</component>`,
}, {
  role: "user",
  content: "a simple todo app",
}, {
  role: "assistant",
  content: "<"
}]

But the answer always starts from an empty string, the followed by “<” and then goes the actual code. And no matter what starting sequence I provide it seems to be ignored or included in the answer.
I use streaming and model “gpt-4o”

2

You cannot provide an assistant message as a prompt, only as past responses as a history.

Chat Completions already adds an unseen assistant prompt, and it does not have a method to continue upon text you provide.

1 Like
3

Hmm, is there a way to extend the output context?
I was using this approach with Anthropic to extend the output context limit - I would append the previous answer as an “assistant” answer in a chain of calls if the answer was stopped during to context limitation

4

OpenAI has not offered continuation within an assistant response. Anthropic apparently knows that there are several uses for such a feature and allows further exploration and capabilities.

1 Like