AI reels are not made by an AI video generator

Ask people how to make AI reels and most will describe a video generator: type a prompt, wait a minute, download a clip. That clip is the part of the job that matters least. The generator does not know what your business does, cannot write a hook, has no voice, no captions, no publishing schedule and no idea whether yesterday's reel worked.

So what does a working AI reel pipeline actually look like, which stages need an AI model at all, and where does a video generator belong in it? Here is the pipeline Ultim runs today, with the numbers from our own reels.

Stage 1: the brief comes from the website, not from a prompt

The first thing the system does is read the site. It collects the facts, the numbers, the people, the tone, the photos (up to 24 per brand), the screenshots and any video it finds on the pages (up to 6). Every fact in the brief carries the quote it came from.

This is the rule that makes the rest safe: if a claim is not on the site, the writer is not allowed to use it. The reel can say "676 shops" only if the site says "676 shops". For a business this matters more than any visual choice, because a reel that invents a price or a result is a liability, not marketing.

Stage 2: the script is written hook first

The writer gets the brief and produces a script: a hook of 3 to 8 words that is on screen at frame zero, 6 to 8 beats of 2 to 4 seconds each, a caption and an end card. The hook follows the Veritasium rule we use everywhere: break a belief first, ask the question before giving the answer.

Length is not a matter of taste. Based on our analysis of 4.5 billion TikTok posts (2014 to 2026), videos of 61 to 120 seconds appear 3.08 times more often among top 1% posts than average, and 8 to 15 second clips appear at 0.63. A caption of 21 to 100 characters with no hashtags lifts 1.7 to 1.8; an empty caption scores 0.27. The script writer applies these as defaults, so nobody has to remember them.

A deterministic QA step then checks the script in code: every on-screen headline must appear in the spoken line, every number must exist in the brief, the end card must be last. Scripts that fail are rewritten, up to three times.

Stage 3: the voice is recorded, with timestamps

The voice track is ElevenLabs, in any of 29 languages, with word-level timestamps. At $0.00022 per character a 296 character script costs about $0.07. The timestamps are what let the on-screen words light up as they are spoken, and what let the editor cut scenes on the voice rather than on a fixed grid.

The data gives a small reason to prefer a synthetic voice over a borrowed song: own voice or own recording as the sound carries a lift of 1.09, someone else's music 0.79. An AI voiceover is original audio on your account.

Stage 4: the scenes come from real material

This is the stage people assume needs a generator. In our production style it does not. Kinetic, the style every Ultim account gets today, is motion design built from the brand's own facts, numbers, logo, palette, photos and screenshots: type that lights up as it is spoken, counters, bar races, a photo pushed in on the beat. No AI images, no AI video.

The reason is quality, not cost. The owner's rules for the style, written after rating dozens of reels, include "real material only: site photos, screenshots, logo, palette. No illustrated or fake UI, no AI people." Generated pictures of a product look like generated pictures of a product.

Where a generator does belong is the talking-person format. For a UGC-style reel, a still of a believable person in the business's world is generated, then animated into 5 second talking clips by an image-to-video model with native speech. We tested six of those models for $2.47; the table and the pick are in the model comparison.

Stage 5: the edit is code, not a timeline

Kinetic reels are rendered with HyperFrames, HeyGen's open-source engine that turns HTML, CSS and GSAP animation into an MP4 through headless Chrome and FFmpeg. One HTML document per reel, scenes stacked on a timeline, scene starts snapped to the nearest music beat within 120 ms, music ducked about 18 dB under speech, two to four subtle sound effects, loudness normalised to -14 LUFS. Output 1080 by 1920 at 30 frames per second.

On the production worker a render takes 82 to 122 seconds. We wrote up the whole method in how Claude makes motion design.

Stage 6: publishing is the boring part done right

The reel goes to TikTok, Instagram Reels and YouTube Shorts through the official APIs, in the afternoon in the account's local time (10:00 to 18:00 local carries a lift of 1.04 to 1.12, with the peak at 15:00 to 16:00), with the AI-generated flag set on all three platforms. You can approve each reel, change a sentence (only that part is re-recorded and re-rendered), or switch on auto-publish.

Stage 7: the learning loop is the stage nobody else has

Each reel tests one variable: the hook, the length, the format, the hour. After 24 and 48 hours the system reads views, likes, shares and saves. A format that fails three times is benched for a week; a format that gets shared gets a second slot; the writer avoids angles you skipped.

Why 24 and 48 hours? In our own 100-reel experiment on a new TikTok account, almost every reel got its views in the first 12 to 24 hours and then went flat. There was no slow build. The median reel got 303 views, the best 1,728, and the shape of the curve was the same for all of them. The verdict is fast, so the loop can be fast.

What each stage costs

Stage	What does the work	Cost per reel (Oct 2026)
Brief and script	Language model, deterministic QA	$0.10 to $0.20
Voice	ElevenLabs, $0.00022 per character	$0.05 to $0.09
Scenes and edit (Kinetic)	HyperFrames on a render worker	under $0.01
Scenes (UGC talking clips)	Image model plus image-to-video with speech	$0.62 to $1.15
Publishing and measurement	Platform APIs, scheduler	near zero

The Kinetic numbers are from five production reels made on 13 October 2026: $0.155 to $0.291 each, $1.009 in total. The UGC range is the production estimate for a 15 to 21 second talking reel. A full breakdown is in what one AI reel really costs.

What this means for your reels

If you are choosing a tool, ask where it starts. A tool that starts at "type a prompt" skips the brief, so everything downstream is generic. A tool that starts at your website has facts to work with, and facts are what hooks are made of.

Spend your attention on the hook and the learning loop, not on the generator. The metadata levers (length, caption, sound, hour) are worth about a 3x multiplier at best; the first three seconds and whether you keep posting decide the rest. Based on our analysis, an account with 100 or more posts has a 39% chance of at least one top 1% hit; one that stops at four has 1.6%. The pipeline exists so that reaching 100 is a scheduling detail rather than a heroic effort.

Ultim runs all seven stages from one pasted link. The free first step is the hook: paste your site into /tools/reel-hooks and see whether the five hooks it writes from your own pages are worth a hundred attempts.

FAQ

Do I need an AI video generator to make AI reels?

Not for most business reels. The motion design style renders real photos, screenshots and numbers with no generated images at all. A generator is useful when you need a person talking to camera, and then only for that scene.

How long does it take to make an AI reel?

In Ultim the script write takes a few minutes, the voice seconds, and the render 82 to 122 seconds on the production worker. The slow part is the learning loop: results are read at 24 and 48 hours before the next reel is planned.

How much does one AI reel cost to produce?

On 13 October 2026 five production motion-design reels cost $0.155 to $0.291 each in API spend. A talking-person reel built from generated clips costs $0.62 to $1.15.