How one app got 55 short videos in an afternoon

On 1 November 2026 we pointed Season Reels at a tailoring shop app in the morning and came back about four and a half hours later to 55 reviewed vertical videos across 11 storytelling lanes. Nobody opened an editor. Nobody wrote a script by hand. The first person to look at the work saw finished renders and five-frame contact sheets.

This post is the run as it happened: what ran in what order, what it cost, and what went wrong. The numbers are from that run, not a projection.

The order of work

Survey, ten minutes. Before anything creative, the orchestrator collected the brand assets (mark, wordmark, colours, fonts), the existing marketing, the locale files with the real strings the app shows in English and Hindi, and the shape of the machine: cores, memory, ffmpeg, which processes must not be disturbed. Everything downstream depends on this being right.

Two briefs in parallel, twelve minutes. A research agent with web search wrote the research brief: formats that work for the audience, pain points in the audience’s own words, the seasonal calendar for the next twelve months with posting windows, licensed audio sources with their exact terms, and twenty one-line hooks. A second, read-only agent wrote the facts sheet from the code and docs: every feature, exact prices and limits, a must-not-claim list, and forty to sixty real interface strings per language. The facts sheet is the guardrail for everything that follows.

Foundation, thirty minutes. Two more agents in parallel. One built the design kit: safe zones, fonts that render Devanagari properly, headline and caption components, a counter that groups digits the Indian way, device frames, and app screens rebuilt from the real strings. It had to produce one complete sample video and fix what its stills showed before handing over. The other built the audio library: music cuts across eight moods, normalised to the same loudness and trimmed to whole bars, twenty-four synthesised sound effects, a brand sting, and a text-to-speech pipeline that emits word timings for captions.

Lanes in two waves, about 25 minutes each. Eleven lane agents, six in the first wave and five in the second. Each got the shared rules and five concept sketches for its lane, naming the hook, the mechanic, the look and the music mood. Agents could swap a concept for a better one of the same spirit. The rules that mattered: at most two phone-led videos per lane, three distinct visual systems per lane, read the other lanes’ notes first so nothing overlaps, make stills, look at them, fix, then render.

Review, continuous. For every finished lane, the orchestrator regenerated the contact sheets itself, looked at all five frames of every video, and judged them against an acceptance list. Precise redo notes went back to the same agent, which still had its context. A redo round took three to seven minutes. We review from stills on purpose, and after two failed rounds a concept is replaced rather than polished.

Finish, fifteen minutes. A full typecheck, a probe of every file for resolution and audio, and a loudness sweep. The set landed between -18 and -21.5 LUFS integrated, which is where short-form platforms want it.

What went wrong

The first wave surfaced four problems, and the fixes went into the second wave’s prompts.

• The lower half was empty. Several videos ended their content a third of the way down the safe area because agents read “the platform’s interface covers the bottom” as “stay high”. The fix was to state composition targets as pixel ranges rather than adjectives, and to say “vary the vertical layout” in every prompt.
• Text inside phones was unreadable. A phone frame 400 pixels wide puts app text at 14 to 18 pixels. The fix: a 24 pixel minimum for body text on the canvas, and wider phones whenever the screen carries the message.
• Caption stickers covered buttons. A sticker on top of a mock interface reads as a bug. The fix was a caption band in clear space, or captions beside the device.
• Contact sheets went stale. The stills tool sometimes reused an old build, so an agent’s sheet did not match its render. The orchestrator now regenerates sheets before judging, and when in doubt pulls frames from the video file itself.

None of these were visible to the agent that made the video. All of them were visible in five frames.

What it proved

Three things carried the run, and they are the three things Season Reels is built around.

Honesty came from the facts sheet, not from review. Lane agents cited it when they changed a brief: the app had no “duplicate order” button, so that concept went; there was no documented setup time, so a video said “five steps” instead of “five minutes”.

Variety came from mechanics, not topics. Each lane was a different way of telling, and reading the other lanes’ notes removed almost all overlap. The one near miss, two stopwatch videos in different lanes, was caught by the agents themselves.

Parallel agents never shared a file. Eleven agents edited the same project for two hours with no conflicts, because each owned its folder and a single registry imported fixed file names.

A second run, for a mobile game, made three gameplay videos and a store trailer from fresh captures. That is the whole track record so far, and we would rather you knew it than guessed.