How to make viral short-form video: the mechanics, not the luck.

A viral short is engineered, not lucky. The mechanics are documented: hook the first three seconds, hold retention to the last frame, close a loop, caption every word, give viewers a reason to send it, and post often enough to keep entering the lottery. The platforms have confirmed most of the ranking signals on the record; the most-watched creators teach the rest. This tutorial collects both, with ten named examples and the AI pipeline for producing shorts at the cadence virality demands. It sits in the Library beside the rest of our guides.

What makes a short-form video go viral?

A video goes viral when its per-viewer metrics beat comparable posts in its first test batch. Every major platform now works the same way: a new post is shown to a small audience, its completion rate, rewatch rate, and share rate are measured against similar content, and the winners are promoted to a larger batch. Virality is that promotion loop repeating. The test batch is served largely to strangers, so it does not depend on your follower count, production budget, or luck; it measures what a sample of strangers did in the first hours.

That is good news for anyone willing to treat it as engineering. Each stage of the loop maps to a craft decision you control: the hook decides whether the test batch watches at all, retention decides whether they finish, the loop decides whether they watch twice, captions decide whether the platform understands what it is ranking, and share triggers decide whether the batch expands itself. The rest of this tutorial takes those stages in order.

The first three seconds decide everything

The first three seconds decide whether the rest of the video gets watched. TikTok’s own advertising research found that 63% of the highest-clickthrough videos state their hook within the first three seconds, and YouTube Shorts’ first ranking metric is viewed versus swiped: the share of people who did not immediately swipe away, visible per video in YouTube Studio. Lose the open and no other decision in the edit gets a chance to matter.

The strongest openings stack three hooks in the same first second, one per channel of attention:

• Spoken hook. The first sentence out of a mouth or a voiceover, and it is the premise, not a greeting. No “hey guys,” no context, no logo.
• Text-overlay hook. The same claim, or a sharper one, on screen as text, for the majority watching muted.
• Visual hook. Something moving or unexplained in frame one: the finished dish, the impossible object, the mid-action moment.

Four hook types cover most viral shorts. Outcome-first: show or state the ending, then explain it. Contrarian claim: open by contradicting what the viewer believes. Curiosity gap: name what is about to happen without showing it. Direct callout: name the viewer (“if you edit video for a living, stop scrolling”). Write five to ten hook variants per video; the hook is the cheapest thing to iterate and the most expensive thing to get wrong.

Retention is the product

Retention, not likes, is what the ranking systems buy. On TikTok, completion and rewatch outweigh raw likes: a short video finished by 80% of viewers beats a longer one with high likes and 16% completion. On YouTube, Todd Beaupre, YouTube’s Director of Growth and Discovery, said in 2025 that watch time is weighted by satisfaction signals, surveys and return visits, not raw minutes, and that Shorts and long-form are ranked by separate models.

The practical consequence: edit backwards from the retention graph. The leaked MrBeast production memo treats the graph as the edit’s source of truth; a line like “we lost 21 million viewers in the first minute” described above-average retention at that scale. Every second has to earn the next one. Cut the sentence that does not, even when it is your favorite. And note Beaupre’s satisfaction weighting: a viewer who finishes, feels tricked, and never returns is worth less than one who finishes and comes back. Retention tricks that cheat the viewer get discounted downstream.

What does each platform reward?

The three feeds rank on different first metrics, and the differences are documented. Instagram head Adam Mosseri confirmed in January 2025 that the three dominant Reels signals are watch time, likes per reach, and sends per reach. TikTok weighs completion and rewatch above everything. YouTube Shorts starts with viewed versus swiped, then satisfaction-weighted watch time. On all three, first-hours test-batch performance decides scale.

Sends per reach deserves the emphasis. Likes are a tap; a send is a viewer staking social credit on your video with a named person. It is the hardest signal to fake and the one Instagram singled out, which is why the share-trigger section below treats it as a design target rather than a byproduct.

Loops and pattern interrupts

A loop is an ending that lands on the first line, so the video restarts before the viewer notices it ended. Rewatches count as watch time twice, which makes the loop the highest-leverage structural trick in short-form. Zach King’s “Magic Broomstick,” the most-viewed TikTok ever, is a seamless visual loop; Jenny Hoyos ends every Short on a twist that references her opening sentence, so the restart feels like a reveal.

Inside the video, pattern interrupts do the same job at smaller scale: a cut, an angle change, a prop, a zoom, a new location, roughly every two to four seconds, so the viewer’s attention never settles long enough to swipe. PJ Accetturo’s rule for the Kalshi ad was a new spectacle every two seconds. Alex Hormozi’s editors cut on every sentence. You do not need that density for every format, and the Sam Sulek example below is the proof, but you need a reason each second for the thumb to stay still.

Captions and social SEO

Keywords in speech and on-screen text now do the classification work hashtags used to do. Platforms run speech recognition and OCR on every upload, so the words you say and show are how the system decides who sees the video; hashtags have been demoted to a minor signal, with TikTok search the partial exception. Use three to five tags at most and spend the effort on saying the topic out loud in the first line instead. The trade even has a name: social SEO replaced hashtag strategy.

Captions are the other half. Roughly 92% of feed video plays muted, a number from an older Verizon Media study that practitioners still plan around, so a video without on-screen text is mute to most of its audience. Word-by-word captions add a second input channel: the viewer reads and hears the same sentence, which lifts both comprehension and retention. Hormozi’s ALL-CAPS keyword-highlighted style became the default look of the genre. Caption every word either way; restraint is a styling choice, not an excuse to skip them.

What makes people share a video?

People share what makes them feel something strongly and look good for sending it. Jonah Berger’s research at Wharton, collected in Contagious, found that high-arousal emotions such as awe, anger, and amusement drive sharing, while low-arousal states such as contentment and sadness suppress it. His STEPPS framework names the levers: social currency, triggers, emotion, public visibility, practical value, and stories.

To engineer sends rather than hope for them: give the video one named recipient (“send this to the friend who still color-grades by eye”), make it an identity claim the sender endorses by forwarding, or pack it with practical value dense enough that forwarding it is a favor. A video that makes the sender look informed, funny, or thoughtful gets sent; a video that only makes its creator look good does not.

How often should you post?

More often than feels comfortable, with flattening returns. Buffer’s October 2025 study of 11.4 million TikTok posts across 150,000 accounts found that accounts posting 2 to 5 times a week earned 17% more views per post than once-a-week accounts; 6 to 10 times a week earned 29% more; 11 or more earned 34% more. Per-post performance rises with volume, then flattens.

Cadence also feeds the test-batch lottery: every post is another draw, and the draws teach you. Ten videos a week produce ten retention graphs to learn from; one polished video a month produces one. The creators in the examples below post daily or near-daily, and the AI pipeline at the end of this tutorial exists to make that cadence survivable for a small team.

The frameworks the best creators teach

Five frameworks cover most of what the most-watched short-form creators teach, and they compose: hook by one, structure by another, end on a loop.

• Hormozi’s Hook-Retain-Reward. From $100M Leads: earn attention, hold it with open questions and movement, then pay it off so completely that watching felt like a good trade. The reward step is the one amateurs skip.
• Hoyos’ outcome-first loop. Hook with the outcome, foreshadow the path, connect beats with but/therefore, end on a twist that lands back on the first line. Built for 90% retention, and measured against it.
• MrBeast’s three-act retention arc. Front-load the thumbnail’s promise in the first minute, escalate through the middle, save a payoff for the end. The edit is built backwards from the retention graph, not forwards from the script.
• The but/therefore spine. The South Park writers’ rule: if two beats connect with “and then,” cut one. Every beat should force the next through “but” or “therefore.” It is the cheapest retention tool that exists, because it costs a script pass and nothing else.
• Give the ending first. Show the finished dish, the transformed room, the final score, then rewind. Curiosity about how replaces uncertainty about whether the video is worth finishing.

The sixth framework is the thesis of this tutorial: platform-native camouflage. Every viral AI-made format borrowed an existing native grammar: the Bigfoot wave borrowed the GoPro vlog, AI ASMR borrowed satisfaction loops, the Kalshi ad borrowed the sports-book commercial, the brainrot memes borrowed the TTS meme. The “look what AI made” era is over; nobody watches capability demos twice. A format-native premise, executed with AI speed, is the durable skill.

Worked examples: ten videos and why they traveled

Ten named examples, each reduced to the mechanic that made it work. Between them they cover every principle above.

1. MrBeast’s production memo. The leaked “How to Succeed at MrBeast Production” memo (September 2024) shows the operation from inside: videos edited backwards from the retention graph, the thumbnail’s promise front-loaded, and losing 21 million viewers in the first minute counted as above-average retention. Why it matters: the biggest channel on YouTube treats retention as the entire product.
2. Jenny Hoyos. Roughly 600 million Shorts views in her first year, targeting 90% retention. Her formula, laid out on Creator Science episode 167: outcome-first hook, foreshadow, but/therefore beats, twist ending that loops to the first line. Why it matters: the loop formula is teachable and she teaches it.
3. Alex Hormozi. Word-by-word ALL-CAPS captions, keyword highlights, zero intro, a cut on every sentence. Why it matters: he industrialized caption-first editing so thoroughly that the style now signals “ad,” and 2026 taste is correcting lighter. Copy the discipline, not the look.
4. Zach King, “Magic Broomstick.” The most-viewed TikTok ever, 2.2 billion views, Guinness-certified. Fully visual premise, seamless loop, no language barrier. Why it matters: the ceiling belongs to videos that need no translation and no ending.
5. PJ Accetturo’s Kalshi ad. An NBA Finals spot made solo in 2 to 3 days for about $2,000: Gemini and ChatGPT for the shot list, Veo 3 for generation, 300 to 400 clips cut down to 15, a new spectacle every two seconds. Over 3 million views on Kalshi’s X in a week, and publishing the prompts went viral on its own. Why it matters: overgeneration is the workflow, and the breakdown outperforms the artifact.
6. Neural Viz. Josh Kerrigan’s one-person sci-fi comedy universe, around 238,000 YouTube subscribers, built on Runway Act-One, Sora, Kling, Veo, ElevenLabs, Suno, and Premiere. Why it matters: persistent characters and writing carry the channel; AI is the production department, not the premise.
7. The Bigfoot vlog wave. May to June 2025: the first Bigfoot selfie-vlog pulled 6.3 million plays in three weeks. Why it matters: the format borrowed GoPro vlog grammar for an impossible subject. Native grammar, impossible content: the camouflage thesis in one clip.
8. AI ASMR glass fruit. #AIASMR was reported at 640 million views within roughly 90 days of appearing, built on impossible satisfaction: glass strawberries sliced cleanly, perfect loops, 8-second completions. Why it matters: complete-and-rewatch is the whole design; the format is retention math wearing a lab coat.
9. Sam Sulek. From zero to roughly 3.5 million YouTube subscribers in a year with unedited 20 to 60 minute gym vlogs, about six uploads a week. Why it matters: the counterexample. Satisfaction can come from intimacy instead of cut density; the cadence rule still held.
10. Italian brainrot memes. Billions of cumulative views in 2025 from a pipeline of ChatGPT image generation plus ElevenLabs TTS. Why it matters: virality came from remixability, not polish. The format invited a thousand collaborators; the originals were almost beside the point.

The AI pipeline for shorts

An AI shorts pipeline turns one person into a daily-cadence studio: an LLM drafts the script and five to ten hook variants, a video model generates the shots, and editing tools handle captions and cuts. For generation, the working models are Veo, Runway, and Kling; our model comparison covers picking a model and the platform comparison covers picking a tool to run it in. Plan for 20 to 25x overgeneration: the Kalshi ad generated 300 to 400 clips to keep 15, and that ratio is typical, not wasteful.

Around the picture: voice with ElevenLabs, music with Suno, long-to-short clipping with Opus Clip or Descript, captions and the final cut in CapCut. If you are prompting the video model yourself, the craft transfers directly from our cinematic prompting tutorial, and the short-film pipeline walkthrough covers the same stack at narrative length.

One warning, and it is the section thesis restated as policy: platforms have moved against undifferentiated mass-produced AI content, most explicitly YouTube, whose 2025 monetization update targets inauthentic mass-produced channels by name. AI in the pipeline is rewarded, because it buys cadence and iteration. AI as the product is penalized, because a feed full of identical capability demos is a feed people close. Everything upstream of this section, hooks, retention, loops, sends, still decides the outcome; the pipeline just lets you attempt it daily.

Resources

Free first. Everything below is either free or a book worth its price.

• The MrBeast production memo: the retention playbook of the biggest channel on YouTube, from inside.
• Colin and Samir’s YouTube Playbook: the creator-economy analysts’ distillation of what works.
• Creator Science: Jay Clouse’s podcast and newsletter; the Jenny Hoyos episode alone is a course in Shorts.
• The Publish Press: creator-economy news three times a week.
• Buffer’s research blog: large-sample cadence and format studies, including the 11.4M-post TikTok study above.
• vidIQ’s blog: working documentation of YouTube’s ranking signals.
• Alex Hormozi, $100M Leads: the Hook-Retain-Reward chapter is the useful part for this tutorial.
• Jonah Berger, Contagious: the share-psychology research behind STEPPS.
• PJ Ace on X: publishes full prompt breakdowns of every AI ad.
• Know Your Meme: trend archaeology; the fastest way to study a format wave after the fact.
• Richard van der Blom’s annual LinkedIn Algorithm Report, for the posts side of this craft: see the sibling tutorial on viral social posts.

Questions creatives ask

What makes a short-form video go viral? Winning the test batch. Every platform shows a new post to a small audience, measures completion, rewatch, and shares against comparable posts, and promotes the winners to a larger batch. Virality is that loop repeating, and every stage of it maps to a craft decision you control.

How important are the first three seconds? Decisive. TikTok’s advertising research found 63% of the highest-clickthrough videos hook within three seconds, and YouTube Shorts ranks first on viewed versus swiped. Stack a spoken hook, a text-overlay hook, and a visual hook into the same first second.

What does each platform reward? TikTok: completion and rewatch over raw likes. Instagram Reels: watch time, likes per reach, and sends per reach, per Mosseri, with sends the one to engineer for. YouTube: satisfaction-weighted watch time, with Shorts and long-form on separate models. All of them: first-hours test performance.

How often should you post? Buffer’s 11.4-million-post study: 2 to 5 posts a week earns 17% more views per post than one, 6 to 10 earns 29% more, 11-plus earns 34% more, flattening as it climbs. Cadence compounds because every post is another test batch and another retention graph to learn from.

Can AI-generated video go viral? Yes, when AI is the production department and the premise is format-native: the Kalshi ad, the Bigfoot vlogs, Neural Viz. Platforms suppress undifferentiated mass-produced AI content, so AI as the product is penalized while AI in the pipeline is rewarded.