Allow injecting system messages in chat

Feature Request: On the Open Assistant website, when on the chat with the assistant, allow for a single "Assistant" message to be injected before the user input.

Rationale:

The model gets additional training via the thumbs up and thumbs down buttons and we use this to score the responses. However, there are cases that even after 10 retries and fiddling with parameters, the model is incapable of producing a correct output. Therefore all answers are downvoted. I don't think this is very useful without a correct answer to compare against.

However, it is possible to "inject" hints that would make it produce a correct output. As an example, here's the user input I gave:

HINT: flat_map
NOTE: Reply in a concise manner, write tiny snippets. Make your answer AS SHORT AS POSSIBLE.
In rust suppose we have an iterator  (.into_iter()) over a Vec<String> we want to collect into a new vec. I want something similar to a map() function, but that instead of yielding a single element, it could yield many. This is because I might encounter an item that is too big and I might want to split it into smaller pieces, creating more elements in place. An iterator should be able to do this, but I don't know if there's a function to do this.

This creates the correct answer after editing. However, voting thumbs up for this does include the prompt engineering and the training might be not as effective as it could be.

What I am suggesting is to have an extra option to inject these hints before or after my user message, possibly marked as "assistant" for the tool. These hints can be associated with the current message.

This would allow on the data you collect to differentiate what are hints and what is the actual user request.

I really don't care how this is implemented in the UI. It could be even a button that injects a marker in the textbox such as "<----- PLACE YOUR ASSISTANT HINTS BELOW ---->", and the tool can parse it afterwards. Or a textbox that pops up. All that matters is that this option is visible in the UI in such a way that everyone uses it in the same way.

No files or tests are named. Start by tracing the Open Assistant website's chat UI and message submission flow, then determine how messages are represented and stored; done means users have a visible, consistent way to add one assistant hint associated with their message and the collected data distinguishes the hint from the request.