Build a better prompt for general scientific writing

[j6k4m8]

Right now, the LLM will sometimes respond with a trailing "@user: i did it!" or whatever, sometimes it'll strip the newline, sometimes it'll dup the whole line... A more specific prompt, and more guidance might do the trick.

[wrongu]

I'll work on version 0.1 prompts, porting over what I already have in langchain

[j6k4m8]

Awesome — assign yourself here for tracking purposes?

[wrongu]

Wishlist of assistant functions (not all of these are feasible):

• [ ] bullets --> draft prose
• [ ] bullets or prose --> critique
• [ ] improve readability and engagement per paragraph
• [ ] improve transitions between paragraphs and sections
• [ ] check or suggest references
• [ ] CCC style improvements to 1 paragraph at a time
• [ ] CCC style improvements to abstract
• [ ] detect and fix parallelism violations
• [ ] split paragraphs that contain multiple main ideas
• [ ] various tools for "linking" paper with outline
  • [ ] extract outline from paper
  • [ ] critique outline
  • [ ] outline --> abstract
  • [ ] outline --> title
  • [ ] change outline -> propose changes in paper
• [ ] "review" the paper

[j6k4m8]

i imagine that some of these will be most easily done by having some sort of:

class agent
    can_answer(user_request):
        return llm("does this look like a request to convert bullets to prose?", context, user_request)

    do_edit(user_request):
        ...

[wrongu]

Yeah that would be cool but sounds like step 2 to me, where step 1 is to implement them in a way that works when triggered "manually"

[j6k4m8]

+1!

On Thu, Jun 22, 2023 at 3:03 PM wrongu @.***> wrote:

Yeah that would be cool but sounds like step 2 to me, where step 1 is to implement them in a way that works when triggered "manually"

— Reply to this email directly, view it on GitHub https://github.com/KordingLab/llm4papers/issues/4#issuecomment-1603174068, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAFJKBY4S36FRHACXQZT7I3XMSJGVANCNFSM6AAAAAAY4UDZOE . You are receiving this because you authored the thread.Message ID: @.***>