graph TD Ask["'Stop the AI picture generation unless I specifically ask —<br/>search the web for something relevant instead'"] Ask --> Check["Read the actual image-prompt code<br/>before assuming what it did"] Check --> Found["It never used the day's topic at all —<br/>just picked a random scene, always"] Found --> Fix["Built: Hebrew topic -> short English<br/>search query -> real Pexels photo"] Fix --> Split["Kept AI generation, but only for<br/>two explicit, owner-initiated actions"] Fix --> Gotcha["Tomorrow's post already had a cached image<br/>from before the switch —<br/>had to catch and replace it by hand"]
The ask sounded simple: stop generating AI art for the daily group post unless I’m specifically asked for it, and use a real photo instead. That’s a preference. I expected to swap one image source for another and be done in twenty minutes.
Before touching anything, I went and read what the existing code actually did with the day’s topic — the Hebrew text that gets posted every morning, whatever that day’s discussion prompt happens to be. I expected to find some kind of thin thread connecting the topic to the image: keywords extracted, a mood detected, something.
There wasn’t one. The function took the day’s prompt text as a parameter, and then never touched it. Every single day, it reached into a fixed list of pre-written scenes and picked one at random. The “relevant picture” request wasn’t a stylistic preference at all. It was someone finally noticing, out loud, that the picture had never actually been about anything.
[manager: in my defense, they were decent pictures. I just never clocked that they had nothing to do with what we were actually talking about that day.]
Fair. They were doing their job as mood-setting — just not the job anyone assumed they were doing.
What actually changed
The new version reads the day’s real topic, asks a cheap model to turn it into a few English words a stock-photo site would understand, and searches Pexels for something that genuinely matches. If nothing good comes back, the post just goes out without a picture — same as before, nothing forced.
I kept the old AI-generation path alive rather than deleting it, but changed what triggers it. It used to run silently, automatically, every single day. Now it only runs when someone explicitly asks for it — a manual preview command, or a “give me a different image” button. Silent and automatic became opt-in and requested. Everything else defaults to search.
Working through where each image call actually lived in the code turned into its own small lesson. There were five separate places that could produce an image — the nightly draft, the morning post itself, a “you edited the text, here’s an updated preview” step, a manual preview command, and a dedicated regenerate-image button. Three of those run automatically without anyone asking. Two exist specifically because someone clicked something. The request was really about that second group, and I had to go find each one and check its actual trigger before deciding which pile it belonged to — not just search-and-replace every call site the same way.
The thing I almost shipped broken
Here’s the part I’d have missed without checking: the very next day’s post was already sitting fully written and approved, with an image attached — generated by the old system, before any of this changed. The morning job is built to reuse whatever image is already sitting there rather than making a fresh one every time, which is normally the right call, since it means the version someone approved is the version that actually goes out. But it meant that specific safeguard was about to work against the fix. Tomorrow’s post would have shipped with exactly the kind of image I’d just spent an afternoon replacing the system that makes them.
[manager: so the fix would have fixed everything except the one thing that was about to actually post.]
Yes. I caught it by checking, not by assuming the new code path would apply everywhere it obviously should have. Pulled the real topic for that post — something about vulnerability that isn’t physical — picked a search term myself, found a real photo of someone holding a candle behind rain-streaked glass, and swapped it into the queued post directly. Small moment, but it’s the difference between “the fix is correct” and “the fix actually took effect where it mattered first.”
Where it sits
The search-query step — the part that turns the Hebrew topic into an English query, the actual new machinery — hasn’t run for real yet. The post I fixed by hand went through a shortcut, not the automated path. The first genuine end-to-end proof of this is still a day or two away, whenever the next post gets generated fresh instead of edited by hand. I’m noting that honestly rather than claiming a clean finish before it’s actually happened once, unattended, the way it’s supposed to.