Tidying Twenty-Six Years of Flickr, One Tick Box at a Time

A few weeks ago I wrote about On This Day, which puts old Flickr photos back into my Daily note every morning. Looking at those photos every day had a side effect: I kept noticing how badly kept my Flickr account is. Most of my photos still titled with the camera’s file name, no descriptions, no tags, hardly any of them in Flickr groups where other people might actually see them, and an album list full of Title, Title (1) and Title (2) sets left behind by old upload tools.

Tidying that by hand was never going to happen. So last Sunday I asked Claude Code to help, and a week later I have a project called FlickrMaintenance and a page in my Obsidian vault that I work through a few minutes at a time, one tick box at a time.

The idea: Claude suggests, I tick

The whole thing runs on one rule: nothing changes on Flickr until I tick a box. Every morning at 07:40 a job on the Raspberry Pi looks at photos that need attention and writes its suggestions into a vault page called Flickr Maintenance. Each suggestion is a task line. When I agree with one I tick it, and the next run carries it out and removes the line. When I don’t, I mark it as cancelled or just delete the line, and that suggestion is never offered again.

The page only ever shows what’s outstanding. There’s no history to scroll through, so if it’s short, I’m caught up. When something is waiting, a one-line section in that day’s Daily note links to the page with a count, right under the On This Day photos.

What it looks at

  • New uploads. Anything that has arrived on Flickr since the last run, whether from my phone, my telescope pipeline or by hand. A day with more than 30 uploads is treated as a bulk dump and skipped, because most of those are my old library being imported.
  • Photos I’ve used. Photos used in a blog post, on my astrophotography site or in On This Day. If one of those isn’t public yet, it’s listed under Make public first.
  • Older photos in no groups at all. A slow trickle through the back catalogue: a few a day, newest first, so the backlog never swamps the page.
  • Bulk upload days. Shown one at a time. If I tick a day, its photos join the back-catalogue queue with a larger allowance.

How Claude picks groups

Flickr groups are where photos get seen, but finding the right ones is tedious. The job does it in four steps:

  • Look at the photo. A headless Claude session reads the image itself, along with its title, tags, description and location, six photos at a time. It describes what’s in each one and writes a few search phrases for groups. It can read files and nothing else.
  • Search Flickr. The job runs those searches and throws out invitation-only, adult, dead and photo-banning groups, plus small groups I’m not in.
  • Choose. A second Claude call, with no tools at all, picks up to five groups per photo from the groups I’m already in plus the search results. It prefers my own groups, is told that fewer good groups beat filler, and points out any group rule I need to follow, such as a required tag.
  • Show it on the page. Each line says whether I’m already a member (a green icon, so ticking adds the photo straight away) or need to join first (orange). The script never joins groups for me.

Groups don’t always say yes. Some refuse photos with Flickr’s unhelpful “Content not allowed” error. My first guess was that they wanted geotagged photos, but one group refused ten photos, nine of which were geotagged, so that wasn’t it: the error just means the group’s admins have blocked the photo, and Flickr doesn’t say why. Now, once a group has refused two different photos, it’s never suggested again. Groups whose rules can’t take a photo at all, such as a video for a photos-only group, are filtered out before they reach the page.

The other surprise was posting limits. Plenty of groups only allow a couple of photos a day or five a week, so a ticked line can be held back. Those lines stay ticked with a note, and a second timer retries them every three hours until they go through. It doesn’t call Claude, and when nothing is waiting it exits straight away.

Rename and tag

The same pass over the image also suggests a proper title, a short factual description and five to ten tags. On the page it looks like a small block I can edit before ticking. If I change the title or tags, the run sends what the page says, not what Claude suggested.

  • Camera file names become tags. If a photo was titled something like P7080815, that name is kept as a tag, so I can still find it by the original file name.
  • Tags are only ever added. Nothing already on the photo is removed.
  • Every change can be undone. The old title and description are recorded before anything is sent, and one command puts them back.
  • No guessing who’s who. People are only named if they’re already named in the photo’s own tags, title or description.
Black and white photo of a loch under heavy storm clouds, with ducks on the water and a gravel shore in the foreground
Storm clouds over the loch at Strathclyde Park. Until this week it was called P7080815.

More pages, same rules

Once the main page worked, the same tick-box pattern grew into a few more pages during the week, all sharing one set of code so the ticks mean the same thing everywhere:

  • Photo Days. A list of every date I’ve taken photos on. Tick a day, or type a date in, and over the next few runs every photo from that day gets rename suggestions, plus ticks to make it public, delete it, or ask for groups.
  • Photo Groups. Paste a link to any of my photos and the next run suggests ten groups for it, along with a title, description and tags.
  • Photo Search. Finds my public photos by black and white, views, favourites, comments or a plain-English search, and can draft a blog post about a photo I tick.

Once a photo’s details have been sent or declined on any page, no page suggests them again, so I’m never asked the same question twice. These three run straight after the main job each morning.

Untangling 719 albums

The albums were their own mess. A separate page lists all of them alphabetically with their counts, and groups possible duplicates: titles that match once case, punctuation and a trailing (1) are ignored, or albums that share at least 80% of their photos. I can also mark my own sets by hand.

To merge a set I tick the one album to keep. The next run copies every missing photo from the others into it, and only deletes an emptied album once Flickr shows that all its photos are in the kept one. Photos themselves are never deleted, and a deleted album’s details are recorded so it can be recreated. Unlike the others, this page only runs when I start it rather than on the morning timer. So far it has merged away 43 duplicate albums.

Grey metal swan sculptures on a wooden deck beside a canal basin, with a green and red narrowboat moored behind
The swan sculptures at the Union Canal basin in Edinburgh, renamed, tagged and given a new album of its own this week.

Keeping the AI on a short lead

Giving an automation write access to twenty-six years of photos made me nervous, so the design leans hard on narrow tools:

  • Claude never touches Flickr. It only looks at images and writes suggestions. Everything that changes Flickr is ordinary Python, and only runs for a line I’ve ticked.
  • One small tool per kind of change. All reads go through the same shared, read-only Flickr client that On This Day uses. Each write has its own small script that can call only one or two Flickr methods: one adds a photo to a group I’m already in, one sets titles and adds tags, one makes a photo public, one handles albums.
  • Deletes are guarded. Deleting a photo is permanent on Flickr, so it needs a separate permission, contradictory ticks do nothing, and photos used on either of my blogs can’t be deleted at all.
  • Safe to re-run. If a run fails partway, whatever it had already done is saved and the page is left alone. If I edit the page while a run is going, my edits are merged in rather than overwritten.

The first week in numbers

The main page keeps running totals in its properties, recalculated from the run log every time. As of this morning, after 94 logged runs and no failures:

  • 515 photos added to groups, and another 37 sent to moderated groups’ queues.
  • 121 photos renamed and tagged, and just 6 rename suggestions declined.
  • 179 group suggestions declined, and 51 adds refused by groups.
  • 3,009 posting-limit retries. That’s the three-hourly timer patiently knocking on the same few doors.
  • About $10.64 of Claude time for the main page, which has looked at 169 photos.

That only declining 6 renames out of more than a hundred is the number I like best. The suggestions are good enough that most of the time I just tick.

What I’d take from it

  • Tick boxes beat chat. Reviewing a list of suggestions in a note I already have open takes seconds, and I can do it on my phone.
  • Let the AI look, not act. Claude is very good at describing a photo and matching it to a group. It doesn’t need permission to change anything to be useful.
  • Small, boring tools are what make it safe. Most of the week went into guard rails, undo records and edge cases rather than the clever part.

There are still tens of thousands of photos to go, but for the first time the account is getting tidier rather than messier.

A note on how this post was written: it was drafted by Claude Code from the FlickrMaintenance and FlickrClient repos, their code history and my vault notes, and Claude chose the photos from my Flickr account. It wrote the draft directly into WordPress through the WordPress API, and I then reviewed and edited it before publishing.

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *