I made a small technology investment this week – I paid $200 for an annual subscription to Anthropic’s Cowork, which is their Claude Sonnet 4.6 large language model (LLM) with extensions to make it more useful as a personal assistant. I figured it was time to dive into the current state of AI a bit deeper.
So far it has already paid for itself. There’s a tax preparation step I despise, which is going through a year’s worth of multiple credit card statements and allocating expense to the various businesses I still keep running – the Louisville rental property business, the Turo car rental business, and my independent consulting LLC. That takes *hours* and multiple passes through about a hundred pages of small font pages to produce a spreadsheet report of my tax-relevant expenses. It’s really no fun.
So I loaded up all those pages of credit card statements and told Claude to scan through the data and produce a spreadsheet of expenses with categories I defined. That took the AI about 15 minutes. Then I had Claude edit the spreadsheet and allocated those summed expenses across the businesses according to some rules I gave it. Another 15 minutes. I then did a final edit of the spreadsheet by hand, and voila, 8-10 hours of work was reduced to 30 minutes. And during that 30 minutes, I could multitask and read while Claude did the work.
Claude/Cowork was built to be good at computer-centric tasks, and I can confirm that it’s very good. It generated and used Python scripts behind the scenes to accomplish certain tasks that I would have done by hand.
Next, I had Claude create a flyer for me for the N Galt Avenue Spring Fling event, something I volunteered to do as part of the event committee. I have *zero* artistic talent, so using an AI is the only way I’d ever design a flyer with graphics. That was harder for the AI – it’s not as expert at image processing as it is with number crunching. But the result wasn’t bad (shown below). I’m still working with the AI to update the flyer and improve it a bit, but once again I’m able to do something I couldn’t do by myself – hence the name Cowork, I suppose.
The image processing tasks consumed a lot more time and tokens than the financial analysis task. Honestly, I still don’t understand how the LLM platforms define and use tokens, but they’re basically units of work that comprise the basis for pricing and/or metered LLM usage. In this case the LLM had to try different approaches to image editing – some things it tried did not work. It was impressive to see the software get stuck, back up and try something different to achieve a goal. Turns out that jpegs and pdfs are tough for both humans and machines to work with. So hard that I hit the token limit for Claude/Cowork while trying to fine-tune the flyer. The token allocation resets in 5 hours, so I’ll have to wait until then to finish. As a hobbyist, not a big problem. But if this were paid consulting work…nope. I’d have to upgrade to a $100-200 per month subscription.
Breaking news as I’m writing, SCOTUS just made Convict Donny’s dumbass tariffs illegal. That’s a HUGE setback for the administration and a great thing for American consumers. Trump will lose what little mind he has left, so that should be entertaining. For the first time, he’s found a limit to his executive power. Better late than never, right?
<Insert sarcasm font> And I’m sure the government will get busy and quickly refund the hundreds of millions of $125 BILLION dollars incurred by businesses due to the illegal tariffs. That’s damage that will be very hard to undo.
Next up, someone should get the question of Donny’s ability to drop bombs on any country he chooses, for any reason, in front off SCOTUS. We’re about to go to war with Iran for no fucking reason, and Convict Donny has been silent on the whole question of “why are we going to war”. That needs to change.