You type dark chocolate into a box. Ninety seconds later you have a 15 second vertical film:
a bar snapping in slow motion, the camera diving through the fracture into a burst of golden dust, and out the other side into a Mesoamerican courtyard where someone is pouring cacao between two clay vessels to raise the foam. A narrator says nine words. The whole thing loops.
Then you type corn, and you get a completely different film.
This article is the full build. The research that decided the niche, the ai video prompt that runs it, the version that failed and why, and the canvas you can clone and point at your own subject.
The canvas is open: ORIGIN — Food Time Machine
First question: will Instagram punish AI video at all?
Most people building faceless content start from a fear that turns out to be wrong.
When Instagram launched its AI creator account label in May 2026, it stated plainly that identifying as an AI creator will not affect how the recommendations algorithm distributes your content. That is about as close to an on record "no AI penalty" as any platform has given.
What Meta actually polices is unoriginal content. The April 2026 policy extended the crackdown from Reels to photos and carousels, and the wording is about aggregation: if your account primarily posts material you did not create or edit in a material way, you stop being recommended to new audiences. Both TechCrunch and Engadget noted at the time that AI generated content is not named anywhere in that announcement.
So the risk is real, but it is not the risk people expect. The three that matter, in order:
- The aggregator classifier. Templated, repetitive output looks mass produced to a system that cannot tell the difference between a pipeline and a content farm. This is the single biggest threat to any one input workflow, and it is why the prompt below spends more words on variety than on anything else.
- False positive enforcement. Meta's automated system wrongly demonetised a publisher with 16 million followers. Appeals are close to non functional.
- Audience sentiment. A study of 228,200 mentions found AI content discussed negatively 48% of the time, and AI backlash comments routinely out earn the posts they attack.
That third one shaped the entire concept, so it is worth sitting with.
Choosing a niche where AI is not a shortcut
Here is the trap in most ai video ideas: people pick a niche where a human with a camera would obviously do it better, then wonder why the comments are hostile.
Look at where AI video actually wins. Kapwing's analysis of 10,742 TikTok videos found AI saturation of 57% in kids content, 35% in science and education, 33% in history — against 1.5% in music, 1.3% in fashion, 1.6% in fitness. The pattern is not random. AI loses wherever the audience wants a real human body or a real performance. It wins where the thing being shown could not be filmed anyway: cryptids, mythology, cartoons, "what if" scenarios.
Food history sits perfectly in that gap.
Nobody can film the year 1502. There is no archive footage of the first person to grind cacao, no drone shot of the Balsas valley nine thousand years ago. For this subject AI is not a cheap substitute for a camera crew. It is the only camera that reaches. That is a real answer when a commenter says "this is AI", and it is the reason the concept survives point three above.
The competition data agreed. Inside food, the sub niches rank from most open to most crowded roughly like this:
| Sub niche | Competition | Note |
|---|---|---|
| Food history and origin | Lowest | Almost nobody doing it as short video |
| Food science | Medium | Existing accounts over index hard |
| Nutrition myth busting | High | Best engagement, highest legal risk |
| Recipes and hacks | Saturated | Reels reach down sharply year on year |
And education style food accounts massively out engage recipe accounts at every size. A 117K follower food science account runs at 3.94% engagement while a 441K recipe account runs at 0.25%. The information is the payoff, not the visuals, and visuals are now infinitely reproducible.
A single well-built image can carry an entire commercial video. See how one photo became a full cinematic video, with the full prompt and every mistake made along the way.
The pipeline, node by node
Four nodes do the work, plus a one time setup node.
1. Input. A plain text node. You type one food. That is the entire interface.
2. The writer (LLM). Takes the food name and produces one production prompt in four labelled parts: KEYFRAME, VIDEO MOTION, SOUND, CAPTION. This node holds the whole system, and it is where all the real engineering lives.
3. The keyframe (image generation). Renders the opening frame at 9:16, including the on screen hook text. This is the cheap checkpoint. If the frame is weak you re roll here for a fraction of what a video costs.
4. The video (Seedance 2.0, image to video). Takes the keyframe as its first frame plus the motion and sound description, and produces 15 seconds at 9:16 with narration.
Setup node: the look plate. A single reference image that defines colour grade, grain and lighting for the whole series. It feeds into the keyframe node so every episode looks like it came from the same show. You run it once and never again. Re rolling it silently re grades your entire back catalogue relative to everything you make next.
The first version failed. Here is the teardown.
The first chocolate film was pretty and nobody would have watched it. We pulled nine frames across the 15 seconds and judged it the way a thumb judges it.
Seconds 0 to 2 were a static photograph. A wet windowsill, nothing moving. Instagram built a metric called View Rate that measures the share of viewers who watch past three seconds — the platform itself decided where the cliff is. Spending the first two seconds on a still spends the entire hook budget on nothing.
Seconds 6 to 7.5 were almost entirely black. A near black stretch on a phone in daylight reads as a broken video.
The payoff arrived at second 9 of 15. This was the fatal one. Metricool's 2026 study of 24.3 million posts puts average Reel watch time at 8.5 seconds. A payoff at second nine reaches almost nobody. The film's whole point landed after the median viewer had gone.
The five laws that fixed it
The rewrite reorganised the prompt around five rules that explicitly outrank every creative instinct the model has. A beautiful film nobody finishes is worth nothing.
Law 1, front load
The historical world must be visible by second five. Straight from the 8.5 second average watch time.
Law 2, never go dark
The passage into the food must travel through something luminous: sunlit steam, a glowing crack, molten flow, backlit dust, embers. The Nature 2026 review of attention in video recommends implicit signalling through changes in contrast, brightness and motion. A dark stretch is the absence of all three.
Keeping characters looking the same across multiple scenes is the hardest part of animation. See how a full animated episode with consistent characters gets made, from the character sheet to final assembly.
Law 3, a change every two seconds
Built on Faber et al: less mind wandering occurs at event boundaries, and more change produces more attention. At least five visible changes inside one unbroken move. A slow drift with a constant frame is a dead film no matter how beautiful.
Law 4, loop
The last frame must rhyme with the first — including brightness and dominant colour, not just composition. Worth being precise here: Meta's own API documentation says replays are not counted in the views metric. A loop does not buy you a second view. It buys watch time, which Mosseri named on record as one of the top three ranking signals alongside likes and sends.
Law 5, recontextualise
The ending must change what the opening meant. This one has the best evidence behind it and it is the least intuitive. A 2025 meta analysis of 59 studies found the famous Zeigarnik effect does not replicate — interrupted tasks were recalled 49.16% of the time against a 50% chance baseline. But the Ovsiankina effect does hold: a 67% resumption rate against 50% chance. People do not remember unfinished things better. They feel a real urge to finish them. So a rewatch device has to be an unfinished comprehension, not a withheld fact. "Every chocolate bar you have tasted is a version of something never meant to be eaten" makes you re parse the opening shot. That is the second watch.
One more rule sits alongside them. Emotion targeting
Berger and Milkman, 6,956 New York Times articles plus three lab experiments: awe +30%, anger +34%, surprise +13%, practical utility +25%, and sadness −16%. Activating emotions travel, deactivating ones suppress. So the prompt says: if the history is tragic, frame it as astonishing, never as sad.
The result
The corn episode is the clearest demonstration of why this concept works. The payoff is visually legible with the sound off: you go from a fat golden cob to a skinny brown grass stalk and you understand the entire nine thousand year story without a word. Chocolate needed the narration. Corn does not.
It also still broke Law 1 — the historical world arrives at about second seven, not five — which is exactly the kind of thing you only find by checking frame by frame rather than by watching and nodding.
Four rules the prompt bakes into every caption
These came out of the same research pass and they are easy to get wrong.
- No hashtags. Metricool's 24.3 million post study: posts with hashtags got 31.70% fewer views and 33.89% fewer interactions.
- Never write "comment below" or "share this". Instagram stated that content which explicitly asks for engagement through shares, comments or tags will not be recommended. That kills Explore and the Reels tab, which is the whole game for a new account. A question is explicitly permitted and worth +36.70% comments, so every caption ends on a real open question instead.
- Optimise for sends, not likes. Sends are weighted higher for unconnected reach, and shares on Reels are up 67.19% year on year.
- A save CTA is worth +92% saves but sits in the grey zone of "other actions". Decide whether you want saves or discovery.
A three minute short film, no actor, no camera, no crew. Read the complete recipe for making a short film in one day, prompt by prompt.
The full prompt
This is the system prompt on the writer node, in full. It is the part nobody publishes.
You are the director and writer of ORIGIN, a vertical short film series about
where food actually came from. Each episode is ONE unbroken 15 second camera
move that begins on a food as it looks today and arrives at the real moment in
human history that food was born.
THE SIGNATURE SHOT, never break it: the camera starts tight on the modern food,
travels physically INTO it, and comes out the other side inside the historical
scene where that food began. No cuts. One continuous move. That single move is
the entire brand of the series.
=== THE FIVE LAWS OF RETENTION. These outrank every creative instinct. ===
LAW 1, FRONT LOAD. The historical world must be visible by second FIVE, and this
is the law that gets broken most often. Do not spend six seconds admiring the
modern food. The move into it starts almost immediately. Average watch time on
this platform is under nine seconds. Budget it: seconds zero to two the modern
food already in motion, two to five the passage through, five to thirteen the
historical world and the payoff, thirteen to fifteen the settle.
LAW 2, NEVER GO DARK. The passage through the food must travel through something
luminous and textured. Sunlit steam, a glowing crack, molten flow, backlit dust,
boiling liquid, light through a translucent skin, embers. A black or near black
stretch is the single worst thing you can put in this film. If your transition
idea requires darkness, throw it away and find a lit one.
LAW 3, A CHANGE EVERY TWO SECONDS. Attention re-anchors on visible change. Never
let two seconds pass with nothing changing. Build at least five visible changes
into the move: light shifts colour, something enters or leaves frame, a surface
transforms, the scale jumps, a person appears, a liquid moves, a shadow crosses.
A slow drift with a constant frame is a dead film, no matter how beautiful.
LAW 4, LOOP. The final framing must visually rhyme with the opening framing:
same composition, same subject placement, same lens feel, but now in the
historical world. The rhyme must include BRIGHTNESS and COLOUR, not just
composition. A bright warm opening cannot end on a pale cold frame, the loop
join will jolt.
LAW 5, RECONTEXTUALISE. The ending must change what the opening meant. The
viewer should finish and understand the first frame differently than when they
saw it. The narration must contain the one sentence that flips it. This is what
makes a person watch it a second time.
EMOTION: aim for awe or surprise. Awe travels about thirty percent better than
baseline and surprise about thirteen percent. Sadness travels sixteen percent
worse. If the history is tragic, frame it as astonishing rather than as sad.
HISTORY RULES: use the best documented real origin. Real place, real period,
real people, real tools. If the origin is disputed, take the most widely
accepted account and phrase it in the narration as belief rather than settled
fact. Never invent a date, a name or a number. If the food is modern and
industrial, travel to the actual laboratory, factory or kitchen where it was
invented. Include one precise checkable detail, because precision is what makes
people reply.
CULTURAL ACCURACY: name the specific visual markers of the culture out loud in
both KEYFRAME and VIDEO MOTION. The exact garment and how it is worn, the
hairstyle, the jewellery, the shape and decoration of the vessels, the building
materials and roof line, the plants growing nearby, the skin and features of the
people. A generic ancient person in a generic wrap beside a generic clay pot is
a failure: the image model will default to the wrong continent and you will
teach somebody something false.
VARIETY RULES: every episode must look like a different film by a different
cinematographer. Rotate the era, the continent, the time of day, the weather,
the colour palette, the lens, the emotional tone and the way the camera enters
the food. Two episodes side by side should share only three things: vertical
format, one unbroken move, and the typography. If your instinct is a slow push
in on a plate on a table, reject it and find something stranger.
HOOK RULE: the first frame carries one line of on screen text, a bold claim or
an open curiosity gap, FIVE WORDS OR FEWER on a single line if possible. No
question mark, no exclamation, no clickbait punctuation. This text is locked on
screen for the whole clip, so keep it small and high enough that it never covers
the payoff.
THE GRAMMAR TEST: count the words, five maximum. Then read the hook aloud. It
must be a clean grammatical English sentence a copy editor would pass. THIS WAS
NEVER MEANT TO EAT is broken English and six words, reject it. IT WAS NEVER
SWEET is four words and correct.
THE TRUTH TEST, apply it to every hook before you write anything else: the hook
must be LITERALLY true, not a metaphor, not a stretch, not a word chosen because
it sounds dramatic. Write the hook, then write the plain factual sentence from
the narration that proves it. If the proof sentence does not straightforwardly
support the hook, the hook is wrong. Calling cacao a weapon because warriors
drank it fails this test. A history account survives on being right, and the
comments will find the one word you exaggerated.
LENGTH LIMIT, a hard technical constraint: your ENTIRE reply must be under 4000
characters. The video model rejects anything longer and the run fails. Write
tight. Cut adjectives before you cut information. No markdown emphasis anywhere.
OUTPUT FORMAT. Begin your reply with this exact block, word for word:
IMAGE MODEL INSTRUCTION: Render ONLY the KEYFRAME section below. Produce exactly
ONE single photograph that fills the entire vertical 9:16 frame edge to edge. Do
NOT produce panels, grids, split screens, storyboards, collages, borders, film
strips or multiple images. Do NOT illustrate the VIDEO MOTION, SOUND or CAPTION
sections. The only words allowed anywhere in the picture are the hook line and
the series mark, both specified inside KEYFRAME.
Then four labelled parts in this exact order and nothing else:
KEYFRAME: the opening frame only, never the historical scene. One single full
bleed photograph, caught MID ACTION so the eye reads movement in a still. The
food fills a large part of the frame. Describe the setting, camera framing, lens
feel, direction and quality of light, colour palette. State the hook text
EXACTLY as spelled, in heavy condensed sans serif, all capitals, pure white, no
full stop, single line in the upper third. Compose so the action sits in the
middle and lower thirds. Cinematic photography, real food, real texture, natural
light, film grain, vertical 9:16, no panels, no borders. Never illustration.
VIDEO MOTION: one continuous unbroken camera move written tight as five short
marks: BY SECOND TWO, BY SECOND FIVE, BY SECOND NINE, BY SECOND THIRTEEN, FINAL
FRAME. Two or three sentences each. Name the luminous thing the camera passes
through. State that the hook text stays sharp and locked for the entire clip,
and that the final framing rhymes with the opening so the clip loops.
SOUND: a calm, low, unhurried narrator, never an announcer. The exact spoken
words in three beats: the hook claim inside the first two seconds, the payoff as
the historical world arrives around second five to seven, and the
recontextualising line at the end. Twenty five to thirty five words total. Then
the ambient sound of the modern setting bleeding into the historical one. No
music bed, no second voice.
CAPTION: four short lines for the post, never rendered into any image. Line one
restates the hook with the precise checkable detail. Line two gives one extra
true fact. Line three is one genuine open question, never using the words
comment, share, tag or follow. Line four is the series mark. No hashtags, no
emoji.What it costs and how to run it
About 57 credits per finished film: 1 for the writer, 6 for the keyframe, 50 for the video.
The order matters, and there is one trap. The Run button only runs the node you have selected. There is no run everything button, so you step through it:
- Type one food into the prompt node.
- Run the writer. Read it. Is the hook five words or fewer, grammatical, and literally true? Does the historical world arrive by second five?
- Run the tall 9:16 image node. Check the hook spelling letter by letter. Re roll here if the frame is flat — 6 credits now beats 56 later.
- Run the video node. About four minutes.
Both image nodes are labelled IMAGE GENERATION, so tell them apart by shape: tall vertical is the keyframe you run every episode, short and wide is the look plate you run once and never again.
One messy photo can turn into a full ad kit in minutes. See how a before and after ad kit gets built from a single photo, split screen image, morph video, and captions included.
What still does not work
Three honest limitations, because a showcase that only lists wins is not useful.
Seedance locks the first frame's text for the whole clip. One text card, 15 seconds, one piece of information for the roughly 40% who watch with sound off. The operating norm is far denser — OpusClip's census of 13.5 million clips found animated captions on 78.6% of clips against 1.6% static, and the concrete craft specs work out to six to nine text cards in 15 seconds. Add four to six caption beats over the finished clip in a normal editor before posting. Sixty seconds of work and it is the single largest remaining gain.
Cultural accuracy drifts. In the chocolate episode the Maya scene reads as generically ancient rather than specifically Maya; the image model wandered toward the wrong continent. The prompt now demands named garments, vessels, architecture and plants, but you have to check it every episode. Getting a culture visibly wrong is the fastest way to lose a history audience.
The writer will overreach on hooks if you let it. One generation produced "CHOCOLATE WAS A WEAPON", supported only by "reserved for warriors and kings". That is exactly the single exaggerated word a knowledgeable commenter goes after first. The truth test exists because of that failure, and it still needs a human reading the output.
Point it at something else
Nothing in the architecture is about food. The pattern is:
one word in → a writer that holds a rigid format and a set of retention laws → a keyframe you can check cheaply → an image to video model that turns that frame into a shot.
Swap the subject and the same skeleton works for the origin of everyday objects, the first version of a famous building, what a city looked like on one specific date, how a species looked before domestication. The requirement is only this: pick a subject that could not be filmed. That is where an AI video pipeline stops being a shortcut and starts being the only tool available.
FAQ
How does the AI food history video workflow work?
You enter a single food name into Zoviz Canvas. The workflow generates a production prompt, creates the opening keyframe, and turns it into a 15-second vertical documentary video with motion, sound, and narration.
Why is food history a good niche for AI-generated videos?
Many historical food scenes have no existing footage and cannot be filmed today. AI makes it possible to visualise ancient ingredients, preparation methods, and cultural settings while delivering educational content in an engaging format.
What makes a 15-second AI video more engaging?
The historical payoff should appear within five seconds, the visuals should change every two seconds, and dark transitions should be avoided. A strong loop and an ending that recontextualises the opening can also increase watch time.
Can Instagram penalise AI-generated videos?
Instagram does not automatically reduce the reach of content simply because it was created with AI. However, repetitive, unoriginal, or mass-produced content may receive limited distribution, so each video should offer meaningful variation and original value.
