Refactoring the shadowmaker has become a bigger headache than I had originally anticipated, but it’s for the long-term health of the system, so I’m sticking to my guns. This weekend added further drama when I finally stopped running away from frontmatter and embraced it for all my metadata. Sure, scattering #ch-command directives throughout the body of the notes was insane, but fixing it is going to mean more than just adding a few metadata fields. I may have to completely change the way I use Obsidian.
After trying a few different schemes for generating bag names, I’ve made some further decisions about how the system will work. Not as automated as I’d hoped, but not as manual as I’d feared…
◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇
After exploring several different ways to generate candidate bag name phrases, I found that one of the most effective was to scan online image databases and pull title phrases and alt tags from those. They almost always use evocative, visual language and do so in just a few succinct words. Using this source will give me plenty of variety - in terms of grammatical structure, subject material, vocabulary, etc. - and produce much more varied ideas than what I was getting from AI-based phrase generators.
So the phrase source is solved, but as discussed previously, I still need to review them manually. Not just to stop unfortunate phrases from slipping in, but also because the difference between a dull phrase and an evocative one can be very subtle.
For example, during one test, the generated phrase was “cows in fields.” That’s an entirely serviceable phrase to use as a bag name, and it’s a pleasant enough image, but it isn’t much fun. Compare that with the image I got when I changed a single word.
In my view, “cows in boots” creates a much more engaging image, and there’s absolutely no way an algorithm is going to know that. So I’m realizing that I need to add a tenth constraint:
- The bag name must be evocative of a fun, playful image.
I don’t mean they all have to be funny, but prosaic imagery about livestock isn’t really my thing. On the other hand, cows in candy colored rain boots feels very on brand. Which means I not only have to review the bag names for decency - I also have to adjudicate the fun potential.
In the process of doing this, I also discovered that I care more than I thought I would about which image gets selected. Even with a good seed phrase, you still get some duds. Have a look at the candidates generated for the phrase “yellow corn hat” and see if you can spot the keeper among the chaff.
This means my process is going to be even more time consuming than originally planned. As a test, I generated a month’s worth of phrases, edited them, generated candidate images for each, selected a winner from each set, and then loaded the results into the Happy Skunk database. That process took me about an hour. I can probably get that down by maybe 50% once I figure out where the tedious parts are, but even if I can’t, spending one hour per month to keep the community fed with a daily dose of fun seems like a reasonable investment.
Anybody want to play “tuxedo raccoons”? If so, here’s what the opening rack looks like, so you can be sure you typed it in properly.
