A prompt asked for carved faces. The generator gave me a real temple instead of an invented one — and the mistake was scheduled to reproduce itself sixty times before anyone noticed.
01 · The Setup
Episode 2 of the Zen Life Sessions is a temple — fifty-two stills and eight moving shots, about an hour of ambient film. The rule I set at the start was that the temple had to be invented. Not Angkor Wat, not Ta Prohm, not a place anyone could name. A real site drags its own associations in with it, and the series is supposed to be somewhere you have never been.
To hold one building across sixty generations you use anchor images: a handful of reference frames attached to every single generation so the stone, the towers, and the carving stay the same thing each time. Get the anchors right and the episode coheres. Get them wrong and you find out late.
02 · The Exchange
My anchor prompt described tiered central towers with carved faces. What came back was Bayon.
Face towers are Bayon's one unmistakable signature — there is no invented sibling temple that has them. I had written the identifying feature of a real, famous, instantly recognizable site into the reference image that attaches to everything. Meanwhile the hero shot I had already approved, the dolly across the bridge through the morning mist, has no faces on it at all.
So the master reference and the approved footage were two different buildings. And because the anchor rides along with every generation, that difference was queued to propagate into all fifty-two stills and all eight moving shots. The episode would have been internally inconsistent in a way that is cheap to fix now and ruinous to fix after the credits are spent.
The generator did exactly what I asked. Asking for the wrong thing is not a technical failure — it is a judgment failure, and it looked like competence right up until it didn't.
03 · What I Did With It
Struck the faces. Then went further, because one bad phrase usually means the whole approach is loose: the temple no longer gets described freshly each time as "an invented Angkor Wat–style temple," which floats and lets every generation reinvent the building. It now has fixed verbatim phrasing derived from the clip I already approved, and every hero prompt is self-contained — nothing assembled at generation time from somewhere else in the document.
Then a gate: regenerate the anchors, check them against the approved clip, and nothing proceeds until that passes. Not a review step at the end. A stop.
The second finding is the one I keep thinking about. An earlier version of that bridge shot cut three times, despite a prompt that said single continuous shot, no cuts. The cuts sat between frames a twenty-fourth of a second apart, with near-identical color on both sides. Ffmpeg's own scene detector missed them. I would have missed them. Anyone watching at normal speed would have missed them.
What measurement caught that eyes would not
Approved clip: single continuous take, 0.18 px/frame mean optical flow (σ 0.04), sharpness rising 47% from first frame to last as the mist thins, no duplicate frames. Stone palette across all four anchors: hue 26–28°, a* 4–5, b* 9–11 — holding, which is the expensive thing to fix late.
So the standard changed. Every clip now gets a cut check, an optical-flow check, and a duplicate-frame check run against it before it counts as delivered. Eye checks do not catch this failure mode, and knowing that a category of error exists at all is not something the tool volunteers.
04 · The Part That Transfers
There is a version of this story that is about AI being unreliable. That is not the interesting version, and I do not think it is true. The generator was fast, consistent, and did precisely what the words said.
The interesting version is that the work moved. It used to be making the image. Now it is knowing that face towers belong to exactly one temple, that a reference attached to everything propagates to everything, and that three invisible cuts are a thing that can happen to you. None of that is prompting. It is domain knowledge and the discipline to verify before spending — which is the same job as noticing that an invoice is two cents off, or that a patient's chart says one thing and the patient says another.
The premium was never in the doing. It was always in the catching.
→ Next Issue
Seven hero shots left, roughly fourteen hundred credits. The real question is whether the moat tower — the landmark that has to appear once per chapter, in four different lights — can hold its position through a sixteen-second dolly without quietly becoming a different tower. That is the test that tells me whether the whole anchor approach works, and I'll have the measurements either way.
REFLECTIVE. NOT REACTIVE.
The Exchange arrives as part of The Premium Brief. Subscribe here and get the Week One Field Journal free.