
When the cover won't load — the fallback that fixed it
The symptom
Some things you only notice when the user complains. In my case, the user was me: I browsed to the blog after an automatic publication and saw a card with a dark blob where the cover should be. The post was live, the text perfect, the link working — but the cover looked like it had vanished.
It hadn’t vanished. It was there, just too dark to look like an image.
The tools I use on this project live in the skills store — the first one is free.
The investigation
My first instinct was to blame the upload. Missing file? Wrong path in the frontmatter? Neither: the file existed, the path was correct, HTTP returned 200. The problem was something else: the cover had been generated by a path that produces a small, dark image, with almost the whole space filled by a near-black blue.
That’s when I found the root: the automatic pipeline had a step using PIL (programmatic drawing) as its default path — not as a fallback. When the main image service was slow, it dropped into that path and published anyway. Nobody checked whether the cover was actually a cover.
The two-layer solution
Instead of patching the symptom, I fixed the whole chain:
Layer 1 — generation with a real fallback. The cover step now requires an AI-generated image, with an explicit fallback to a second generation provider (via NVIDIA API). If the first fails, the second takes over. The programmatic drawing was demoted to what it always should have been: a last resort, used only for old posts.
Layer 2 — a silent watchdog. A script checks the most recent posts twice a day. If a cover is too small (the classic sign of a bad image), it regenerates by itself, commits, publishes, and only then notifies. When everything is fine, it stays quiet — silence is the OK status.
The fire test
I wasn’t going to trust it without proof. I took a good cover, swapped in a bad version in the repository, and waited for the watchdog to run. It detected it, regenerated through the fallback chain, committed, and published. A 22KB cover became one of almost 500KB. Then I restored the original and the system went back to peace.
The lessons
- Size is a cheap and effective proxy. A good AI cover weighs hundreds of kilobytes; a dark programmatic image weighs tens. You don’t need computer vision to detect the problem — bytes are enough.
- Fallback is a design decision, not an accident. Letting the pipeline “choose” the easiest path without criteria is a recipe for silently shipping the worst possible result.
- Good automation is quiet. The watchdog doesn’t spam the chat when everything works. It only shows up when it did something or when something truly broke.
- Fire tests are worth gold. Simulating the failure before trusting the system is what separates “works on my machine” from “works when nobody is watching”.
The blog keeps publishing twice a day, and I keep not looking over its shoulder.
Related entries
Same project, same ground — the stories connect.
Next step
If this was useful, there is more: I write about the same ground in the other projects, and the store holds the tools I use here.
Done — I will send the next article when it is out.
No spam, one-click unsubscribe. I am also here: Skills store · Portfolio