A detailed document that highlights every repeatable format we cracked, how each one was arrived at, what was locked and what was swappable, and what every single run returned.
We had one message - invideo has all the latest AI models - come use it for anything AI-video related. The goal was to get this message to reach as many creatives as possible while making sure the message didn't get diluted. Virality first all the way.
So rather than explain the product and hope for distribution, we went and found what the target audience was already consuming, saving and sending - and built repeatable formats that delivered the message inside those behaviours.
Three formats came out of it. Each one is a slightly different manifestation of the same underlying idea, and each one was built to be run again and again by swapping a single variable.
people were sent the invideo link directly in their DMs. Every format carried an explicit comment trigger; every comment was answered with a one-pager on that post's content plus the steps to execute it on invideo. That is the size of the pipe running from the feed into the product.
Take a visual world people already want to copy (a popular brand, a music video, a recently released movie). Break down the science of why it looks the way it does. Turn that into a detailed prompt structure that people can use to borrow the visual language for their own projects. Ship it as a carousel. The format remains the same, the visual worlds keep changing.
Run 01 · the visual world of the film Sinners, released at that moment.
Run 02 · a fashion campaign language that was blowing up on Instagram.
Run 03 · an editorial visual world, decomposed into prompt components.
Run 04 · same format, pointed at a just-released album. Timing landed.
The target audience was AI creatives - editors, content creators and AI creators who use AI models and tools on a regular basis to generate images and video. We had a set of these people sitting in the office, so rather than theorise about them, we asked them.
It wasn't easy to get the prompt right to get the exact thing you needed. That was the pain point. So: take really cool visual worlds, decompose them, explain the science of what works in that world, and break it into components you can translate into a prompt.
Shelf life matters. Today the underlying models are good enough that plain language often decodes what you want, so handing over a raw prompt means less. The format was built for its moment - and it read that moment correctly.
Which world you point it at. That is the only thing that moved across four runs.
Every world we picked was live at that moment - Sinners was releasing, Jacquemus was blowing up on Instagram, the Bully album had just dropped. Because decoding what's trending is the format, timing is not a separate bet you have to place. It's built in. When it lands, the sky is the ceiling. When it doesn't, the floor is still substantially high.
Sinners World. The softest of the four and still a substantial result. No intervention was needed - the format was already rock solid.
Kanye / Bully. The album had just come out and we timed it correctly. Same format, same build, nothing else changed. It skyrocketed.
Whenever someone commented, we sent them a prompt doc: a step-by-step instruction doc that walks them through how they can sign up on and implement the exact visual world on invideo.
People still need an AI video platform to actually execute what they just learned. So the comment mechanism is a direct route to send people to try invideo. This runs across any post carrying a comment trigger - the doc always pairs that post's core content with the directions to implement it on the platform.
The reel manifestation of the same principle. Break down the philosophy of a visual style and pair it with where that style is winning right now. The first MVP was good, the second iteration was a home run and did 4.23M views.
Run 01 · the Dune visual world, timed to the new film coming out at that point.
Run 02 · the Nike visual world, decomposed the same way.
Run 03 · an out-of-the-box shot. A visual style that had never been decomposed before, taken because he was blowing up at that moment.
The starting point was a post from a creator called Ohneis. He caters to a design-heavy audience - he is principally a designer - and he breaks down design styles: paper-cut style, collage style, 3D style. Every post carries the philosophy of that design, plus how that design is blowing up right now and how it is getting used across the industry. Philosophy, plus inspiration for usage.
We took that as inspiration and started doing it in our own way - our own breakdowns of the visual styles of popular worlds.
This ran hand in hand with the carousel manifestation. The carousel format is similar in nature; this is the reel manifestation of the same principle. Even inside one format, a small tweak produces a different close fraternity of format - and it can work just as effectively.
Which world you point it at. That is the only thing that moved across three runs.
Each of these reels carried an explicit comment trigger - the Dune run asked people to comment “PROMPT” for the exact system behind Dune-style visual language.
The doc is a one-pager on that post's core content, plus a step-by-step on how to get those results on invideo. People still need an AI video platform to execute what they just learned, so the comment mechanic is a direct route to trying the product. It runs across any post carrying a comment trigger.
The format worked. It just wasn't blowing up to the degree we wanted. And the reason mattered more than the result: getting there would have taken a lot of posts.
On Ohneis's page, every post is that treatment. He has trained the algorithm to find him an audience that values that zone of treatment - that is what makes it work for him. We were not going to get the time to let things compound the same way on our page.
Solid, repeatable, reliable. Three runs, one architecture, consistent floor. Nothing wrong with the format itself.
Compounding takes volume. A page full of one treatment trains the algorithm. We didn't have that many posts to spend.
So the decision was not to drop the format. It was to keep the format and change what sits on top of it.
The format underneath stayed exactly the same: take a visual world, give it a prompt structure people can use to recreate it for whatever they want. What changed was reading what people were actually valuing - and it was the prompt itself.
The real sequence is two beats, not three. First they need to see the visual style is genuinely cool and something they want. Then they want the actual prompt that produced it. The science of the visual world in between doesn't cater to anyone and everyone.
They have to want it first.
The science of the world. Doesn't cater to everyone.
Out in the open, creating FOMO: these are the exact prompts we used.
Bring beat one and beat three together, drop the middle. The creative execution was adapted from a post on another page that had done something similar - “this prompt helped me achieve this visual” - re-cut for our own insight.
Landed right as Kanye's new album came out, so it picked up and skyrocketed - likely a top Instagram post that month. Getting to three or four million views is not something that happens just like that.
Another shot at the same format and the same execution. Also did fairly well.
| Per run, averaged | Original | Evolved | Shift |
|---|---|---|---|
| Views | 36,676 | 2,163,340 | 59× |
| Reach | 22,570 | 1,400,321 | 62× |
| Follows | 55 | 18,893 | 343× |
| Saves | 302 | 38,404 | 127× |
| Shares | 166 | 43,755 | 264× |
| Avg watch time | 10.3s | 14.0s | +36% |
| Skip rate | 59.2% | 43.1% | −16.1pt |
Skip rate is the honest one. It moved before the views did - the prompt-forward cut held people in the first seconds, and the reach followed.
The supply-side format: each run is an ad for a specific feature. Take a top pain point people have with AI video right now, and link it to the feature or workflow inside the product that already solves it. The format never changes - only the pain point does.
Run 01 · pain point: AI visuals fall apart the moment you need continuity - shot one works, shot two drifts, shot three is a different film. Answered by Vision.
Run 02 · pain point: holding one product identical across a whole set of shots.
Run 03 · pain point: output resolution. Answered by UltraRes upscaling - the split-frame before/after carries the whole argument.
Run 04 · pain point: AI skin that reads as plastic. The strongest retention of the four - lowest skip rate, highest comment volume.
This bucket is supply-side. Where the other two formats were built demand-first off what the audience already wanted to consume, this one starts from the product and earns its way to the audience.
The insight is a pairing: take the top pain points people have with AI video right now, and link each one to an existing feature - or an existing workflow inside invideo - that already solves for it. Not a feature announcement. A problem the viewer is currently living with, answered.
Who it's for is the same as the other two formats - it's still catering to the same audience. Only the core insight of what they get is slightly different. Every one of these formats is a slightly different manifestation of the same underlying thing, treatment included.
Which pain point you pick up, and which feature answers it. That is the only thing that moved across four runs.
The same visual philosophy runs through all four posts. Very high cinematic visuals. Unusual composition. A deliberate stylistic choice in how the frame is built and how the post speaks - distinctly European in its visual design and its tone of address.
That consistency is what keeps four separate feature ads reading as one body of work rather than four unrelated product posts. The pain point changes; the treatment holds it together.
High-end film look, not software-demo look.
Framing that doesn't sit where you expect it to.
How the visual design reads, and how the post speaks to people.
Same mechanic as the other formats. The Vision run asked people to comment “vision” and got sent the link.
On this format the route is shorter than on the others, because the thing being explained is the product. The pain point is named, the feature is the answer, and the comment hands them the way in.
These are product-marketing posts that were built to be mirrored as ads and creator briefs. They are not trying to do what the pop-culture formats do, and comparing them on raw reach misreads what they were for.
The interesting movement is in retention. Skip rate tightens steadily across the four runs - 62.4% to 47.6% - and the two best-retained runs are also the two that converted follows.
Tight, predictable band. All four runs landed between 40K and 51K views. A supply-side format with a reliable floor is a rare thing.
Realistic Skin. The lowest skip rate of the four, and 594 comments - the most of any run. Pain point sharpness shows up in retention first.
The two runs that named the most concrete, most currently-felt problems - resolution and skin realism - took 460 of the format's 618 follows between them. The sharper the pain, the better the ad.
Three formats, thirteen posts, one body of work. Strip out the specifics and this is what it actually proved - each principle with the run that demonstrates it.
Sit with the people you're making for. Ask what they actually consume, what they save for later, what they send to friends. Study the pages already winning with them and dissect what those pages do. The format is the output of that exercise - not a guess you validate afterwards.
If more than one thing changes between posts, you don't have a format - you have a series of one-offs. Lock the structure hard enough that the only decision left each time is which world, which style, or which pain point you point it at.
If decoding what's culturally live is the format, timing stops being a separate bet you have to place correctly. Every run is automatically pointed at something people already care about - which is why the ceiling is uncapped and the floor still holds when the timing doesn't land.
The same underlying principle, moved from carousel to reel, is a different close fraternity of format - and it can work just as effectively. You don't need a new insight every time. You need a new manifestation of one that already works.
When something works but doesn't break out, the question isn't what to build instead - it's which beat the audience isn't actually paying for. Cut that, pull what they are paying for to the front, and watch retention move before reach does.
A comment trigger isn't engagement bait when the thing you send back is genuinely the thing they asked for - the post's content as a one-pager, plus how to execute it on the platform. People still need somewhere to actually do the work. That's the whole mechanism.