I remain deeply cynical about much of the hype ~ particularly from our political leaders and Silicon Valley grifters ~ that surrounds GenAI and am worried about the sustainability of it (though this thread from Alex demonstrates what a complicated thing that is to really get your head around.)

…but at the recent Product for the People in London Matt Webb gave a really inspiring talk (which Karina has brilliantly captured) about basically noodling and experimenting with AI at a more personal level – building micro-apps that have no market but are useful to you or just a way to express some creativity.

Now when the web was still open I loved to ‘make’ things with Yahoo Pipes and the like so I decided to take a bit of a break from being Abe yelling at clouds and see if I can learn a bit about how these things really work by building a few little things to help me out.

I signed up for a paid Anthropic account for no other reason than I liked the name Claude 🤣 and gave myself two little projects related to my jobs newsletter.

1. Try and automate my long list that I start with every week but automating the search for a list of job titles (around 30) from the DDaT* framework against a list of jobs boards.
   
2. Generate a list of popular links from my followers on Bluesky so I can share some interesting blogposts etc each week – but also to just help me keep up with what is the geeky discourse without having to be quite so online.

Some meta things I learned. More than half the battle was getting my MacMini set up with everything it needed. Blooming hell, no wonder developers are always grumpy 😆. Getting Python and Node and all the relevant dependencies sorted definitely caused some swearing but given it has been years since I tried anything like this it is not surprising (and never on this hardware). My rustiness with Terminal was also no surprising!

Also as if I didn’t already know it but the web really is not open anymore – not a bit – and these two experiments demonstrating the difference between one of the few corners of the web trying to be open in BlueSky and the commercial world of job boards etc.

So starting with the success. Claude and I eventually identified a way to achieve something resembling my goals for the Bluesky Popular Links generator. I think it ended up being just short of 25 versions back and forth over three hours before I started to ask the right questions – there were a number of false dawns and dead ends – but eventually I ended up with a Node app that started by generating a list of the 100 Bluesky users I interact most with by giving them a score based a bunch of criteria that the Bluesky API provides (this is a bit rough and I suspect not entirely right but it is pretty close), it then looks for URLs they have shared or liked and then generates a list (in Markdown) where more than two of them have done that with the same URL.

I was able to follow bits of the Bluesky API enough to make some specific suggestions and link Claude to bits of the documentation so I felt like I was contributing a bit and learning as I went. My prompts/questions definitely improved as I went along and I was really pleased with the outcome. It was a fun and educational afternoon and I’ll definitely use the little app – and have some ideas for improving it.

That was the peak though. The job search adventure has been more in the tragedy genre. Jobs boards really don’t want people interacting with them in this manner – certainly not for free – and I’ve been running into walls more than I’d recommend! Initially Claude recommended a Python scraper based solution based on my list of job titles and job sites. After all the swearing to get Python and the dependencies in shape (the aforementioned swearing) I finally got it to run…and it was an abject failure as all the sites were like Batfink with his wings like a shield of steel against scrapers.

Various efforts to circumvent this failed ~ though I suspect I wasn’t asking the right questions/prompts here. I broke it down to one job board at a time to simplify things but still no luck.

Next we tried to find job boards with open APIs or endpoints that might be useful. This was a bit of a red herring initially as the suggestions were all in the US to start with (I forgot to remind Claude I was in the UK) and even when the UK was targeted the web is just not very open. There were limited APIs at all let alone ‘open’ ones and even where they were discoverable the results were uninspiring. Amusingly LinkedIn was the most open but the results were pretty broad! I suspect with more patience this might end up being semi useful. The Guardian seemed like it should work but Claude and I ended up giving up and every other site I use regularly was a non starter. Basically I wasted four hours and got nowhere – and also don’t know whether it was the fault of Claude or the modern web…but I actually suspect the latter.

So what did I learn? Well it is fascinating how Claude is able to generate – and debug – code that actually does what you ask it – give or take. I suspect my engineer friends would be horrified by the outputs and I definitely wouldn’t rely on any of it for anything important but for someone like myself with zero development chops but occasional ideas that might be fixed by a bit of code – and who just about knows the right kind of questions and sources to help – I can see that I’ll keep paying for it for the time being at least.

I’m not ready to jump on the hype train but maybe my cynicism has dimmed (a little).