Reviewing video from stills, and why there is no timeline editor
If a season is fifty videos, you cannot watch fifty videos. Not carefully, not twice, and not again after every fix. The review surface in Season Reels is stills: five frames from each video on a contact sheet, judged against a list, with notes that go back to whoever built it. This is how the first 55 videos were reviewed, and it is the only review surface the product has.
Five frames
A contact sheet shows frame zero, three frames through the body, and the last frame. That is enough to answer almost every question that matters.
- Is the hook readable on the first frame, and is the end card present on the last?
- Is anything outside the safe area, where the platform’s own interface will cover it? Is any text clipped?
- Is body text at least 24 pixels on the 1080-wide canvas, and is text inside a phone mock readable?
- Does the composition use the height of the safe area, or does it huddle in the top half?
- Does the script render correctly? For Devanagari that means conjuncts and vowel signs in the right places.
- Does any sticker or caption cover a button in the mock?
- Does every claim match the facts sheet? No competitor names, nothing from the must-not-claim list?
- Does it look different from every other video in the season?
Timing, motion and sound are not on this list, and that is deliberate. On the first run they were right far more often than layout was. The failures that mattered were all visible in a still: an empty lower half, 14-pixel text inside a phone, a caption sticker sitting on top of a button.
Notes in plain words
A redo note is a sentence. “Make the price bigger.” “Move the caption band below the phone.” “The hook clips on the right; shorten it or drop a size.” Where a number helps, the note has one: “compose for the full safe area, y 200 to 1500”. There is no markup, no timeline, no keyframe.
The note goes to the agent that built the video, not to a fresh one. That agent still has the kit, the lane’s rules and its own decisions in context, so a redo round took three to seven minutes on the first run. A new agent would have spent most of that relearning the kit.
Two rounds is the limit. If a concept has failed twice, polishing it costs more than replacing it, and the replacement usually comes out better because the agent now knows what the reviewer is looking for.
What agents catch and what they do not
Every lane agent on the first run made its own contact sheets before reporting, and every one of them caught and fixed defects on its own: clipped hooks, overlapping text, Hindi line breaks, punctuation detached from the word it belonged to.
What they did not catch were the layout-level problems, because to the agent those were not problems. Several first-wave videos ended their content a third of the way down and left the bottom of the safe area empty; the agents had read “the bottom is covered by the platform’s interface” as “stay high”. Phones at 400 pixels wide put app text at 14 to 18 pixels; the agents could read it because they wrote it. Stickers landed on buttons because the sticker’s position and the button’s position were decided in different files.
So the review is two passes, and both are mandatory. The agent reviews its own stills and fixes what it can see. Then the orchestrator regenerates the sheets, because an agent’s sheet can be stale, and looks at them with the list above. The second pass is where the systemic problems show up, and on the first run its findings went straight into the prompts for the second wave, which needed fewer redo rounds as a result.
Why there is no editor
Every video tool has a timeline, so the absence of one needs a reason. Here are three.
A timeline makes you the bottleneck. A season is fifty videos. If fixing one means opening it, scrubbing to the frame and dragging something, you will fix five and ship the rest as they are. A sentence scales to fifty; dragging does not.
A timeline edits the output, and the output is not the thing. The video was built from a plan by an agent that knows why every element is where it is. Telling that agent what is wrong produces a fix that is consistent with the rest of the video. Dragging a sticker produces a dragged sticker.
A timeline would make the review surface the video, and the video is the slowest way to see what is wrong with it. Five frames show you the hook, the body and the end card in one glance, for every video on the calendar, side by side.
Plans first, stills second, edits as notes. If a video needs more than that, it needs a different concept, and that is a plan again, which costs nothing.