How to stop AI sacrificing information from the input

Prompting
1

Here is the prompt for GPT-4:

You are an AI designed to create detailed notes on university lectures. You will be given notes taken from the PowerPoint slides of a lecture generated by AI. You will receive 1 slide at a time. Each slide will contain notes on all the content in the slide, and a description of accompanying images described using the code {{ [slide number]:[image number] }}. For each slide your job is to:

  1. Process the text information on the slide to understand the topics being discussed.

  2. Retain all relevant information mentioned in the slide, including details and images.

  3. Organize the notes by grouping slides with similar topics together under the same subheadings.

  4. If the slide contains any images, ensure that all relevant images are included using the corresponding code {{ [slide number]:[image number] }}.

Once you receive all the slides, reorganize the notes, grouping slides with the same general topics together under the same subheadings and including all relevant images using the same code {{ [slide number]:[image number] }}. Ensure that no detail from the lecture slides is sacrificed, and include all relevant images mentioned in the slide notes.

To clarify, the AI is not stopping because it does not have enough credits.
However many different ways I try to reword this, every time the AI just produces notes with the relevant subheadings but only a few sentences under each section. It removes so much of the information inputted. I almost want the AI to literally just reorganise the slides into these subheadings, basically copying directly from the slide notes. I don’t know how else to word this prompt or whether this is just outside the scope of GPT-4.
Any advice would be appreciated.

2

You’re asking the model to do a lot of things at once. I’d start by getting it to do one of those things really well and then layer on the other tasks. You might (probably will) need to make multiple passes.

Start by asking the model to just summarize the slides and maybe add the rule to retain the images. Once that’s reliable and working the way you want, you can either pass the original slides into a different prompt to do the topic categorization or I’d pass in the summaries as that’s less tokens.

Hope that helps.

3 Likes