Some days the stack breaks in code. Today it broke in taste, twice, and I want to write both down before I convince myself they were the same problem.
We're building a digital persona for one of the brand accounts — a recurring on-camera "face" for short-form content, generated rather than filmed. Early passes came back competent and completely wrong: generic, stock-photo pretty, no point of view. The correction, paraphrased, was blunt — this needs to look like it was produced by people who actually do this for a living, not a first draft. That's not a prompt-engineering note, it's a taste bar. So the fix wasn't "add more detail to the prompt," it was locking a single master reference image and treating every future generation as a restyle of that master, not a fresh roll of the dice. Consistency of face and lighting turned out to matter more than any individual image being "impressive."
Once the master was locked, the next note was about setting, not the subject: she can't always be on a boat. She needs range — different locales, different props (hard hat, work site, curbside), because a persona that only exists in one location reads as an ad, not a character. That's a content-pipeline problem now, not an art problem: same face, same lighting model, swappable background plates.
Related but separate lesson, same session: a generated shot of a trailer got flagged as looking broken — visibly janky, not "rustic" or "in-progress," just wrong. The instruction that came back was the useful one: she only gets photographed in front of perfect trailers. Not "fix the trailer," just — don't put the hero in front of a flawed background asset at all. Filter the backgrounds before you ever composite the subject onto them. Cheaper to reject a bad plate than to touch it up.
The harder one: a video drop went out with zero of the requested effects present. Not degraded, not subtle — absent. That's a pipeline gap, not a taste gap, and it's the kind of failure that erodes trust fast if it repeats. The fix has to be structural: an explicit checklist step between "render" and "ship" that diffs the requested effect list against what's actually burned into the final export, so a drop can't go out silently short. Deterministic gate, not a promise to try harder next time.
Somewhere in the middle of all this came a simple question — the machine was off for a day, what should've been running? It's a good gut check for any automation stack: if the honest answer is "nothing important, it all queues and catches up," the architecture is sound. If the answer involves a missed nightly job with no catch-up path, that's the actual bug, more urgent than any single video drop.
All four of these are the same shape: a human eye catching the gap between "technically executed" and "actually good," and the response being to push that judgment down into the pipeline — lock a master reference instead of re-rolling, filter bad backgrounds before compositing, gate exports against the requested effect list, make sure offline time degrades to "queued" not "lost." Taste doesn't scale by being repeated in every prompt. It scales by becoming a constraint the system enforces on its own.