I asked for a hunter’s pistol to become a rifle. Everything else in the picture stayed exactly the same — same animals, same background, same pose. Clean, one-item edit.
A few days later I asked for two things in the same request: add a conch shell that belongs in the scene, and remove a child who shouldn’t be in it. Half the animals the song’s lyrics called for vanished. Nothing I’d asked to touch.
What Was Actually Happening
These are the illustrations for my daughter’s nursery-rhyme tab — hand-built pictures for songs, some already carrying real details like her own face in a couple of scenes (that story’s here). Song lyrics keep needing small corrections to the pictures as I notice mismatches: an animal missing that the words mention, a pose that doesn’t match what’s being sung.
The pistol-to-rifle edit was a single, narrow ask, and the model treated it that way — it changed the one object and left the rest of the image alone.
The conch-and-child edit asked for two things at once: add one item, remove another. That’s when unrelated animals from the same scene disappeared. Not touched by either request, gone anyway. My guess is that once an edit request has more than one moving part, the model treats more of the image as fair game to regenerate instead of leaving it alone — but I don’t actually know the mechanism, only the pattern I kept running into.
What Fixed It
Splitting the compound request back into single, narrow asks — one object at a time — got me back to clean, minimal edits. When the object was a real thing with a specific shape I couldn’t describe precisely enough in words (real birds mentioned in a lyric, a specific throwing pose), attaching an actual reference photo and asking the model to match it worked better than describing it and hoping.
🗂 Claude.md Rule AI image edits: ask for one change at a time. A single narrow edit (“swap the pistol for a rifle”) tends to leave everything else alone. A compound edit (“add X and remove Y”) risks the model quietly regenerating unrelated parts of the scene. When the target is a specific real object, attach a reference photo instead of describing it.
Time I wasted: two extra rounds of re-editing and re-checking the picture against the song’s lyrics before I isolated the pattern.