Hi David,

I’m intrigued by what you are trying to do with this LLM project. Prompt-engineering is very much a trial and error process - here’s what I got from walking through it a few times. 

To clarify my understanding of your objective here is: 

• Given the 80,000 hours podcast episodes
• Identify the episodes discussing quantitative social sciences, economics and policy
• Within those episodes identify questions brought up in conversation
• Filter those questions by if they are specific and measurable (operationalizable)
• From there, identify the questions that would impact a donor’s decision-making in deciding whether or not to fund further research
• (gleaning from your examples) add a direct quote and link to relevant literature

Does that sound right? If so here, here’s what I got (a few different threads) (top to bottom)

1st attempt: 

• Make a research agenda for one specific episode
• identify donor relevant funding questions

2nd attempt: 

• first identify all the relevant podcasts (but 10 seems to be the limit for any kind of listing) 
• I narrowed it to just one (quant social). 
• I then tried using 80k podcast filter to help narrow it but perplexity says it can’t read the filterable listings of the transcripts.

3rd attempt: 

• I asked it to pull all episodes related to economics (caps it at 10)
• Pull questions from each episode (only does it for 3 episodes)
• Identify questions that are measurable and influence donor decision-making
• Asked it to add quotes but it lost the thread I guess and didn’t connect to the above list of questions. 
• So I tried narrowing it down to one episode and filter through again adding a quote and a link to a piece of related research. 

Curious to know if these results are along the lines of what you are trying to achieve? 

My guess for systematizing this would be defining a prompt flow that gets the results you want and then replicating it across each individual podcast episode. You could use a separate thread to get it to create a list of episodes relevant to topics to narrow the amount you have to search through. 

Long-term it would be cool to see 80k have something like tim ferriss’ blogbot given how much content they have (and the information that is buried in each podcast episode).