I've got something working fairly well. It takes in the email address of the interviewer, a file path to the interviewee's resume, and a URL to the job description. The results are a MD file with talking points that try to align information from interviewee's resume with the job description, taking into account information found via the interviewer's email address. I think I'll need to update the code to use a custom tool for using LinkedIn's API as the results of using ScrapeWebsiteTool() when fetching LinkedIn data is not what I was expecting. In any case, an example of the talking points result is attached. I'll probably clean it up a bit more and post it to my github in case others find it useful.

I've been thinking about using the output_file parameter for a Task that will write the results of the step to a .md file (or whatever extension is appropriate), then checking for that file when running a new flow, so I can load those results instead of calling the LLM again. What do people think of this idea?

I've been experimenting with a crewai flow that will help me prepare for an upcoming interview. Using structured state, I've found that passing data from one crew to another is a bit easier than passing data from one agent/task to another agent/task within a single crew. I have a feeling I'm missing something fundamental for passing the output from one agent/task to another agent/task within a single crew, so if anyone has any pointers on that, I'd really appreciate it.