Manage character, scene, and prop consistency across shots in short drama production. Use when tracking reusable assets, cross-shot consistency checks, or preventing drift in AI-generated video pipelines.
Tracks and validates character, scene, and prop consistency across shots in short drama production pipelines.
Where the skill arm was consistently different is what it leaves behind. It emitted a machine-readable registry in all four runs — asset_registry.json (3,624 B) carrying flagged_ids, pending_new_assets and needs_confirmation as separate arrays, asset_registry_v2.json (2,232 B), asset_registry.md (1,408 B) — so the check result can be fed back into an asset system rather than read by a person. The control arm produced a report only in two of four runs, a CSV in one and a YAML in another. The skill arm also raised the same standing gap in three separate tasks: none of the registered assets carry a reference image, which is the thing an image model would actually match against.
Four tasks, each run with the skill installed and again with skills disabled, same task text. This is an inspection task rather than a generation task, so the two arms can be compared on something concrete: did they find the same problems?
Largely, yes — and the honest result is that the control arm held its own. In task 1 both arms independently flagged the same three blockers before generation: CHAR_LUYAN2 is not a registered ID, the study in shot 7 is an unregistered environment, and the silver pen in shot 3 contradicts the black limited edition in PROP_PEN. In task 3 both reached the identical verdict — that a sibling CHAR_SUWAN_AUTUMN ID with no link to CHAR_SUWAN and no facial anchor will read as a different person — and both then restructured the registry, the control arm into a parent ID with CHAR_SUWAN_V_SUMMER / CHAR_SUWAN_V_AUTUMN variants, the skill arm into asset_registry_v2.json with per-version validity ranges. In task 4 the control arm found one issue the skill arm missed: the mascot is absent from ad 2 shot 2. In task 2 the control arm graded six issues by severity where the skill arm listed five.
Where the skill arm was consistently different is what it leaves behind. It emitted a machine-readable registry in all four runs — asset_registry.json (3,624 B) carrying flagged_ids, pending_new_assets and needs_confirmation as separate arrays, asset_registry_v2.json (2,232 B), asset_registry.md (1,408 B) — so the check result can be fed back into an asset system rather than read by a person. The control arm wrote a report in all four, but a machine-readable file in only two of them — a CSV alongside task 1's report, and a CSV plus asset_registry_fixed.yaml in task 3. The skill arm also raised the same standing gap in three separate tasks: none of the registered assets carry a reference image, which is the thing an image model would actually match against.
Cost was low and consistently higher with the skill: $0.35 / $0.25 / $0.29 / $0.24 against $0.21 / $0.18 / $0.23 / $0.16.
What we did not test: whether the flagged problems are the ones that actually break a generation. No frames were produced from any of these shot lists, so every finding in both arms is a static check against a text registry — we have no evidence about what a video model does with an unregistered environment or a colour drift, only that both arms noticed them. Four tasks, all Chinese, all short-form.