I have started a S*bst*ck, in which I catalogue my work to hand off the operations of my business to AI. I acquire and operate small e-commerce brands, and this post is about how I've distilled my analysis of prospective acquisitions into one giant prompt.
@SurrealistShip I have a running list of these in a Substack draft for some week when I have nothing else to post. Best one is that it diagnosed my dog’s chicken allergy after his vet repeatedly failed to. Today I added that it diagnosed my a/c issue and walked me through a fix.
I’ve been trying out GPT-5.6 Sol and the new desktop app over the last couple of days and am very surprised to find that it is much more of a strict rule follower than Claude.
Tried to build a tool that scrapes job boards for new listings for my wife, but it proactively checked each board’s TOS and refused the ones that prohibit automated scraping.
Today I have it working on a project where I’m going to do some cold email outreach. Was going to have it pick up where Claude left off and get a Resend account set up to handle outbound, but it refused there because Resend’s terms explicitly prohibit cold emails. That one is fine with me — plenty of email tools and I’d rather use one that’s okay with my use case — but the amusing thing is Claude is the one that suggested it in the first place.
@tszzl@threepointone That’s true but three sizes of model with three thinking levels is too many. If even a user who is extremely well-versed in this stuff can’t clearly know which option is best, it’s too many options and makes the experience worse with zero real benefit.
All this talk about Sol mass deleting files and I’m seriously considering removing the block on rm for Claude because it is a good AI and only ever uses it to delete temp files.
This week's post is about how to best pick the model, thinking level and harness for the task at hand. You'd be surprised how different your results can be giving the exact same task to Opus in Claude Code vs. Claude on the web.
Achievement unlocked: I have two hours left until my Claude limits reset and want to burn through my Fable usage, so I had it spin up so many subagents that it froze my PC.
@TomWoodstom Wow I think this is my most liked post ever somehow. I believe that means I’m supposed to shill my S*bstack about using AI to manage my business, a fund that acquires and operates small ecomm brands: https://t.co/QNCYlXnL52
@TomWoodstom I know I’m never going to need a coax again, but I also understand that is causally determined by the fact that I have a coax in my garage.
@wolfestar4@TomWoodstom Come on now, you can’t just post that without explaining to me what on earth you could possibly have needed two coax cables for…
@p_millerd That seems less like a zirp thing and more reflective of the fact that you can now command a computer, “Implement scripts to try to retain customers who want to cancel. Make no mistakes.”
@sporadica Wrong. It’s Sacramento. Only that and champagne have an ergonomically correct curve on the handle, and Sacramento is clearly the better of the two.
Today's post is about unleashing Claude to do SEO. Am I adding more slop to the internet? I am. Sorry. But a gazillion SEO shops were doing that before AI anyway, and I'd rather sell more products than sit on my high horse.
It seems like some queries are trivially routable to cheap models (e.g. factual questions). Always seemed odd to me to get rid of the whole concept. No routers means people who leave it on instant think AI is much worse than it is, and people who leave it on thinking burn tokens wastefully on simple queries. There ought to be some gains to be had from at least directing the obviously easy stuff to cheap models and letting everything else go to the good ones.