Manual Matt – more misadventures with AI

I’ve had a longer than expected gap between gigs combined with a brush with Covid (for the nth time) so like every other product person on the planet I’ve been experimenting a little with “AI”.

I mentioned in an earlier post I was going to see if I could work with ‘Claude’ to build a Bluesky bot that would randomly post a random daily graffiti/street-art pic from my extensive archives to a new account at a @graffiti.photos (a domain I own but have never used.)

Matthew pointed me at Russs’s Polybot which does most of the hard-lifting and I got Claude to build me a total dummy guide to the steps needed (this is the first incarnation of that step-by-step guide – there were MANY more versions). Now I see a lot of people talking about how they spin things up in 20 minutes using AI. My complete lack of a clue meant it was probably more like 20 hours 😂 but in my defence I did have a bunch of foundational stuff to do.

For a start all my pics are in Google Photos which is pretty much a walled garden these days – all pre-existing APIs are gone now. So I dusted off my Flickr account and upgraded to Pro for a year before migrating my collection over. Sort of. Something went pear shaped and only about 2000 of the 7.5k images made the trip but I decided that was enough to get started and Future Matt could sort the rest out.

Bless Flickr for still supporting so many open APIs by the way – I know it is a bit of a relic but I did appreciate the commitment to the old, open ways of the web. Even if it is due to neglect as much as strategy!

Claude had recommended using a RaspberryPi as the always-on machine to run the bot – but I didn’t have one of those hanging around. I did however have an abandoned MacMini that was my main machine at the height of the lockdown era but had since been neglected.

I got it fired up and then did the upgrade dance – for the OS and Python.

There was some swearing but in the end I had a working machine and bit by bit I pieced together the various permissions and API keys I needed and updated the various Python files and typed things into Terminal with blind trust before eventually (eventually is doing some heavy lifting here) my ‘burner’ Bluesky account displayed a random piece of street-art with a little bit of metadata.

Woohoo. Success.

Except.

You will never see it.

In parallel to all this I’d been ‘Wizard of Oz-ing’ the account manually for a couple of weeks (on https://bsky.app/profile/graffitiphotos.bsky.social and https://www.instagram.com/iamjukesie/ ). Just trawling through my archive and picking an image (trying a different city each day) and posting it manually at a different time each day, adding some different metadata and just seeing what worked.

A few things happened during this experimental phase.

1. I realised that the thing I enjoyed most was exploring my archive and picking a picture rather than just randomly selecting something it almost immediately became much more thoughtful and curated.

2. I started to edit the images a little before posting them – nothing extensive but just tidying them up and fixing the lighting etc.

3. I began to try and identify the artists where possible – something I’ve never really done – which led to following some new accounts on Instagram.

4. Absolutely nobody was/is interested. If anything I have lost followers on Instagram and made next to no impact on Bluesky in three weeks.

5. I have no actual interest in automating it at all. Realising I could was journey enough but I am going to stick to being ‘Manual Matt’.

Really it is a very me thing. I’m just not interested in ‘optimisation’ or ‘automating’ things. My newsletter remains painfully manual as well. I found it a useful exercise and I can see why these tools appeal to people building out little bits of personal, folk software and I suspect I will play around with other ideas in the future just to stave off the slide into irrelevance a little but I guess I’ll continue to be old school – dinosaur that I am!

Leave a Reply