I’ve been using Claude (generally via Claude Code, but also via the Desktop and iOS apps) pretty regularly, trying out different use cases. These are all on the Pro Plan because I don’t wanna spend $100/mo. (Or, really, I don’t think it would be worth it, in my current life.)
Coding, in my experience, feels like a pretty solved use case. Nearly anything I’ve tried I’ve been able to get Claude to succeed at. In some cases, with little input from me to get 80% of the way there. In other cases, there’s a lot more trial and error to get what I want, but it’s still pretty darn good. I could likely get that success rate up using some more tools/MCPs to help it see what it’s building (usually it’s UX stuff that’s the gotcha) more easily. Still, if this is your common use case, it’s going to go pretty well.
Bespoke data analysis has been good, but maybe not quite as good. A while back I took my Netflix ratings data (which I’d saved like 2000 ratings) and uploaded it to Letterboxd. I then exported that data plus more recent ratings from Letterboxd, and uploaded it as a CSV to Claude. I told it to analyze my movie tastes, and then I’ve asked it to make recommendations in certain areas on movies I haven’t seen yet. That’s been surprisingly successful. It’s recommended a few movies I’ve really enjoyed (that I’d not heard of, or had forgotten about), and the recommendations have been pretty reliable. This is probably limited only by Claude’s data set recency and lack of which streaming service a movie might be on (but it tries). I did a similar exercise by exporting my Apple Music/iTunes Library into Claude. This one was ok. I think the amount of data was probably a bit too overwhelming, took a lot more finagling to get it to fit the variety of limits, and then it started making some pretty mundane recommendations (“You like Jenny Lewis? Have you heard of Rilo Kiley?”) and I think it was just too much context to maintain. This is another place that having an MCP or the like directly into Apple Music probably would have worked more effectively to keep all that data from polluting the context window on every request. More on that in a bit.
In this same bucket, I have these lists of top songs of 2025 from some music sites I follow (my friend who posts at Midnight Snark, local music blogger If It’s Too Loud, and internet legend Said the Gramophone). Usually, I’ll find a time I want to check out new music, pull up a page I’ve saved, listen to some. Go back later, forget what I listened to, rinse repeat. Eventually, I give up and the posts stay in my saved archive for a few years.
What a great job for AI, right? Go to those sites, grab the songs, normalize and rank them, and create an Apple Music playlist for me. It has not gone quite as well.
First, Claude tried to access those sites. Of the 3, could only access 1 over the web to pull the data. It tried a bunch of random things (spawning a browser via a tool/MCP was one of them), but was being stymied and blocked. All the while, just burning through tokens and continually needing to compact the convo (losing the context of what it had already done). Then I said “cool, let me just throw these into PDFs for you, then you can read those”. I do that, and now I’m over the context window for the chat I’m in.
I switch to Claude Cowork, point it at the directory the PDFs are in, and it says “thanks, let’s do it”, and then fails ostensibly because it couldn’t install the tool to read PDFs. Except, that’s not the actual reason. When you dig in to that particular set of logs, it’s actually because I hit my limit 😬
We’ll see how this does when I pick it back up again later tonight, but this is where the UX for this process is still too complicated for most people. Managing context windows, managing tool failures, trying 3 or 4 different ways to massage things into the box to get the right outcome. It’s early, and many of these rough edges are going to get smoothed away, but I still think the more complex topics are trickier than something your average user would do.
I also wonder if people sometimes aren’t even realizing that what they asked the computer to do didn’t actually get done. Instead, they set it off and it said “cool” and people have been trained over decades to expect the computer will do the thing they said, and aren’t trained to assume the computer may just completely botch the job.
Still, the pace of advancement is pretty remarkable, and the barrier to solve personal data/technical challenges is basically non-existent now, if you’re persistent and can afford $20/mo.