5 Comments
User's avatar
Gary Mishuris, CFA's avatar

Great article, thank you. Curious:

1) How would you describe the difference in quality of output between your in-depth prompts and much shorter versions?

2) Have you experimented with whether it is best to use the same LLM for both the Master and the "meta" prompt or is there some value to mixing and matching?

Expand full comment
Compound With AI's avatar

It depends on the business complexity. For simple ones i use the standard prompts they handle about 80% of the work.

For more complex ones, it’s better to use the more detailed versions.

For question 2:

I prefer using ChatGPT’s thinking models to improve prompts the quality is generally higher than with other tools. Then I use Gemini to run the DeepSearch.

All of this is based on my personal experience, not on any scientific evaluation

Expand full comment
Ron Cohen's avatar

Thank you very much for writing!

Actually I'm trying to do this (using Perplexity) and I can't.. It doesn't give me any file with many pages, it just answers in a few lines.

First I copied the prompt from step 1 and activated the "deep research" and then I copied, for example, step 2 and it just didn't do anything LOL

Expand full comment
Compound With AI's avatar

I always use Gemini for DeepSearch:

-Open Gemini and select DeepSearch.

-Paste the prompt (not the one from step 1, that one is meant to be used in a standard ChatGPT chat to improve other prompts).

-Gemini will then propose a plan — click Approve.

-The model will build the report and cite sources as it goes. It will take some time.

-Export the result as PDF or Word.

Expand full comment
Ron Cohen's avatar

I didn't understand then what prompt do you use at the beginning?

Expand full comment