kling-4.org writer: continuity log for AI video

The starting problem is a familiar one for anyone assembling AI-generated footage: two clips can each look appealing and still refuse to belong in the same sequence. The article lists the usual culprits — a changing sky, a different camera height, a subject that feels larger in one frame — and argues that adding more descriptive words is not always the useful next step.

The proposed remedy is organisational rather than technical. Before writing the next request, you decide what should stay consistent, expressed as three visual anchors such as warm late-afternoon light, a restrained blue-and-sand palette and a calm camera. Anchors are meant to be specific enough to inspect: "Beautiful atmosphere" is hard to compare, while soft light from the left without a dramatic change in weather gives a clearer review question. Alongside that, each shot gets one job — the wide shot introduces the place, the bicycle brings the scene closer to human scale, the cup provides a quiet ending — so constants and per-shot intent stay separated.

The rest is record-keeping. The article recommends writing down what the interface actually says rather than reconstructing it from memory later, and adding a new row when a shot is revised so you can tell whether a change in framing came from a different reference, a different instruction, or another setting. It is explicit that such a record does not establish causation. The method's value seems likely to hinge on whether the operator actually keeps the log consistent across revisions, and for a three-shot experiment the cost is low — three rows of notes and a text document.