I wanted a way to perform market analysis easier, I just didn't have a way to do so fast, well I tried using the Gpt and Claude chat interface, but then I have to keep track of everything it spits out, and I have to genuinely prompt in such a detailed manner everytime for it to yield good results, because prompting is a skill whether we like it or not, the LLM's can't just assume what we want or need and get it correctly everytime.
Anyway, I wanted to think of a way to sort of speed up the prompting phase but as I wasn't sure how I came by Jules, Now Jules is a google model that works off my requests rather than tokens. So it has 100 requests per day, and in each request it could produce thousands of lines of code or text.
So then I constructed a very detailed prompt and sent it into Jules with different topics I wanted to market research, could be literally anything you want to investigate. But one thing hit the ceiling, I had to keep changing the prompt around everytime so I still had not really "automated" or saved "a lot" of time.
Then I got an idea of getting another agent to instruct Jules for me, so I just brought in an API, now at the time and still, NVIDIA offers free gpu compute models, so I just threw in one of the models and wired it up to nemotron 120b(you could also use any other api), the orignal prompt is very detailed and man-made, but nemotron sort of changes it around depending on the subject to be explored.
So I ended up with an orchestrator, and now all I do is interact with my orchestrator what subjects I want to research, send it in and the LLM will parse it and delegate it to Jules for me, all I have to do is review the pull requests made in a different repo.
Now for those of you who have made it this far reading, this can be seen as "slop generation" although I would like to point out that it is only slop if you let it be, if you construct a good prompt you will get a good result, and the output doesn't have to be taken to the fullest, you could also take it as inspiration or a lead to what you actually want.
The point of this is just to speed up different types of research, whether it is a litterature reivew, market analysis, research paper,experiment, MvP, Proof of concept - you name it.
1 comment
[ 5.5 ms ] story [ 162 ms ] threadAnyway, I wanted to think of a way to sort of speed up the prompting phase but as I wasn't sure how I came by Jules, Now Jules is a google model that works off my requests rather than tokens. So it has 100 requests per day, and in each request it could produce thousands of lines of code or text.
So then I constructed a very detailed prompt and sent it into Jules with different topics I wanted to market research, could be literally anything you want to investigate. But one thing hit the ceiling, I had to keep changing the prompt around everytime so I still had not really "automated" or saved "a lot" of time.
Then I got an idea of getting another agent to instruct Jules for me, so I just brought in an API, now at the time and still, NVIDIA offers free gpu compute models, so I just threw in one of the models and wired it up to nemotron 120b(you could also use any other api), the orignal prompt is very detailed and man-made, but nemotron sort of changes it around depending on the subject to be explored.
So I ended up with an orchestrator, and now all I do is interact with my orchestrator what subjects I want to research, send it in and the LLM will parse it and delegate it to Jules for me, all I have to do is review the pull requests made in a different repo.
Now for those of you who have made it this far reading, this can be seen as "slop generation" although I would like to point out that it is only slop if you let it be, if you construct a good prompt you will get a good result, and the output doesn't have to be taken to the fullest, you could also take it as inspiration or a lead to what you actually want.
The point of this is just to speed up different types of research, whether it is a litterature reivew, market analysis, research paper,experiment, MvP, Proof of concept - you name it.
It is also open-source and available here: https://github.com/Yoosseph/jules-research-factory