[idopmstuff]

I have always found writing documentation to be incredibly helpful for clarifying my thinking. It prevents me from doing mental hand-waving around details, and often times writing down a process that I have done a thousand times is the thing that makes me realize how I can cut steps or improve it.

I'm now in the process of trying to hand off chunks of the work I do to run my business to AI (both to save time but also just as my very broad, practical eval). It really is all about documentation. I buy small e-commerce brands, and they're simple enough that current SOTA models have more than enough intelligence to take a first pass at listings + financials to determine whether I should take a call with the seller. To make that work, though, I've got a prompt that's currently at six pages that is just every single thing I look when evaluating a business codified.

Using that has really convinced me that people are overrating the importance of intelligence in LLMs in terms of driving real economic value. Most work is like my evaluations - it requires intelligence, but there's a ceiling to how much you need. Someone with 150 IQ points wouldn't do any better at this task than someone with 100 IQ points.

Instead, I think what's going to drive actual change is the scaffolding that lets LLMs take on increasing numbers of tasks. My big issue right now is that I have to go to the listing page for a business that's for sale, screenshot the page, download the files, upload that all to ChatGPT and then give it the prompt. I'm still waiting for a web browsing agent that can handle all of that for me, so I can automate the full flow and just get an analysis of each listing sent to me without having to do anything.

[grvdrm]

Question: could you use something like (example) Selenium to perform some or all of those pre-LLM tasks?

[idopmstuff]

I could (I mean in theory - practically, I'm not technically proficient enough to do so), and in fact one of the most promising web browsing agents I've tested is director.ai, which just writes Stagehand code on the fly to achieve the objectives you give it. Unfortunately it can't be invoked via API yet, so doesn't work for my use case.

Honestly, it takes such a relatively small amount of time that it makes sense to just do it myself until there's an agent that can easily handle it; I'm really only spending time trying to automate it now as a test of AI capabilities. If I actually wanted to get it automated tomorrow, the most time-efficient way to do that would just be to involve a VA from somewhere cheap for the work I'm doing.