A film that survives a 5-second skip button, a muted autoplay, a 9:16 crop, and a completion-rate ranking is a genuinely different film from one built for a dark cinema — not a different export of the same one. Platform has to be decided before the story is written, not after the edit is locked. This page covers the universal craft that holds across every destination; each subpage below carries that one platform's actual frame, duration, hook window, and technical spec.
Classical 30-second TV structure spends 5-8 seconds establishing normality, because the viewer cannot leave. That same setup is a dead opening on a platform with an exit button.
1. Cold open. Start after the inciting incident, not before it — the first three seconds have to answer what world this is, why it matters now, and what the image even is. Never open on setup. 2. Front-loaded payoff. Move the best image out of the ending and into the first two seconds, then earn the right to explain it. Research on early-second viewer dropoff (see Sources) found viewers watching under 3 seconds still delivered a meaningful share of total campaign value — if the payoff sits at second 25, most of the paid value never reaches it. 3. Start in the middle (in medias res). Open on the most anomalous frame of the film and let the audience assemble context from what follows. Cheaper than a cold open, since it's a reorder, not a new shot. 4. Hook-Retain-Reward. The dominant short-form frame: a 0-3 second hook (the anomaly), then protect every transition after it with a mini-hook — a 5-second stretch with no cut, no push-in, no new text, and no sound event is a retention leak — and a completion payoff at the end, without which the viewer feels used and doesn't share. 5. The loop ending. End on a frame that flows back into the opening frame, so an involuntary replay generates the rewatch signal short-form platforms weight most heavily. Match the last frame's composition and subject position to frame one; end on an unresolved gesture; never fade to black. This is a short-form-only device — the same loop reads as a mistake on CTV or in a cinema.
Cropping a widescreen frame to 9:16 keeps roughly one-third of the original image — everything below breaks as a result, and needs a different solve than "crop tighter":
The professional pattern isn't "make the 60-second cut and trim it down." It's plan the 6-second idea first, and build the 60 so the 6 already exists inside it.
The order
Decide the hero format (the one with the hardest constraint) → Write the 6-second idea first — one image, one turn; if there's no 6-second version, the concept has no idea, only a story → Write the hero length outward from that 6-second core → Plan every ratio you need from the start, not as an afterthought → Shoot/generate with both ratios in mind → Cut the master at the longest length and hero ratio, locked → Derive cutdowns in order: 30 → 15 → 6, each re-cut with its own structure — never derive the 6 by trimming the 30 → Re-grade and re-mix per delivery; loudness targets differ by an order of magnitude between broadcast and social, so one mix cannot cover both.
Meaning must survive muting on Meta, LinkedIn, X, Snapchat, and any website hero — sound is a bonus there, not a channel.
- Write the plot to be legible from the picture cut alone. Test: mute it and hand it to someone who hasn't read the brief. If they can't say what happened, it's a radio ad with pictures attached.
- Never put a story beat only in dialogue. If a line carries the turn, the turn must also be visible — a reaction, an object, a door.
- Burn captions in, with real typography. Burned-in captions with deliberate design outperform platform auto-captions on completion rate, survive re-sharing and downloading, and don't depend on the viewer enabling anything.
- Design silent-first, then add sound — don't design for sound and hope captions patch the gap. TikTok and YouTube are real exceptions to the mute-first rule (see the platform comparison below); everywhere else, silent-first is the default.
Platform gets decided at two different moments, and getting the order wrong is one of the most expensive mistakes available:
- The logo-sting cold open — two seconds of animated logo is two seconds of the only window a hard-exit platform gave the film.
- Fade in from black on a scroll feed — showing the algorithm a black frame during the exact moment it measures stop-rate.
- The 17-second setup — beautiful on a big screen, invisible everywhere else.
- Letterbox or pillarbox bars — explicitly discouraged by major platforms and read as lazy on any social surface.
- One mix for every platform — a broadcast-loudness mix is inaudible on TikTok; a social-loudness mix fails broadcast QC outright.
Built from official sources
- Structural adaptations for hard-exit platforms, vertical-frame craft, and attention research drawn from vertical-video cinematography guides, published retention/attention studies (Dentsu, Nielsen, Magna, Kantar, System1), and platform creative-best-practice documentation — individual figures carry their own confidence tags on the subpages below where they're cited directly.
The setup tax
Classical 30-second TV structure spends 5-8 seconds establishing normality, because the viewer cannot leave. That same setup is a dead opening on a platform with an exit button.
Five documented structural adaptations for a hard-exit platform
1. Cold open. Start after the inciting incident, not before it — the first three seconds have to answer what world this is, why it matters now, and what the image even is. Never open on setup. 2. Front-loaded payoff. Move the best image out of the ending and into the first two seconds, then earn the right to explain it. Research on early-second viewer dropoff (see Sources) found viewers watching under 3 seconds still delivered a meaningful share of total campaign value — if the payoff sits at second 25, most of the paid value never reaches it. 3. Start in the middle (in medias res). Open on the most anomalous frame of the film and let the audience assemble context from what follows. Cheaper than a cold open, since it's a reorder, not a new shot. 4. Hook-Retain-Reward. The dominant short-form frame: a 0-3 second hook (the anomaly), then protect every transition after it with a mini-hook — a 5-second stretch with no cut, no push-in, no new text, and no sound event is a retention leak — and a completion payoff at the end, without which the viewer feels used and doesn't share. 5. The loop ending. End on a frame that flows back into the opening frame, so an involuntary replay generates the rewatch signal short-form platforms weight most heavily. Match the last frame's composition and subject position to frame one; end on an unresolved gesture; never fade to black. This is a short-form-only device — the same loop reads as a mistake on CTV or in a cinema.
Vertical is not a crop
Cropping a widescreen frame to 9:16 keeps roughly one-third of the original image — everything below breaks as a result, and needs a different solve than "crop tighter":
Master + cutdowns — the professional order
The professional pattern isn't "make the 60-second cut and trim it down." It's plan the 6-second idea first, and build the 60 so the 6 already exists inside it.
The order
Decide the hero format (the one with the hardest constraint) → Write the 6-second idea first — one image, one turn; if there's no 6-second version, the concept has no idea, only a story → Write the hero length outward from that 6-second core → Plan every ratio you need from the start, not as an afterthought → Shoot/generate with both ratios in mind → Cut the master at the longest length and hero ratio, locked → Derive cutdowns in order: 30 → 15 → 6, each re-cut with its own structure — never derive the 6 by trimming the 30 → Re-grade and re-mix per delivery; loudness targets differ by an order of magnitude between broadcast and social, so one mix cannot cover both.
Sound-off design rules
Meaning must survive muting on Meta, LinkedIn, X, Snapchat, and any website hero — sound is a bonus there, not a channel.
- Write the plot to be legible from the picture cut alone. Test: mute it and hand it to someone who hasn't read the brief. If they can't say what happened, it's a radio ad with pictures attached.
- Never put a story beat only in dialogue. If a line carries the turn, the turn must also be visible — a reaction, an object, a door.
- Burn captions in, with real typography. Burned-in captions with deliberate design outperform platform auto-captions on completion rate, survive re-sharing and downloading, and don't depend on the viewer enabling anything.
- Design silent-first, then add sound — don't design for sound and hope captions patch the gap. TikTok and YouTube are real exceptions to the mute-first rule (see the platform comparison below); everywhere else, silent-first is the default.
Decide platform twice — where it enters the pipeline
Platform gets decided at two different moments, and getting the order wrong is one of the most expensive mistakes available:
Universal anti-patterns
- The logo-sting cold open — two seconds of animated logo is two seconds of the only window a hard-exit platform gave the film.
- Fade in from black on a scroll feed — showing the algorithm a black frame during the exact moment it measures stop-rate.
- The 17-second setup — beautiful on a big screen, invisible everywhere else.
- Letterbox or pillarbox bars — explicitly discouraged by major platforms and read as lazy on any social surface.
- One mix for every platform — a broadcast-loudness mix is inaudible on TikTok; a social-loudness mix fails broadcast QC outright.
Built from official sources
- Structural adaptations for hard-exit platforms, vertical-frame craft, and attention research drawn from vertical-video cinematography guides, published retention/attention studies (Dentsu, Nielsen, Magna, Kantar, System1), and platform creative-best-practice documentation — individual figures carry their own confidence tags on the subpages below where they're cited directly.
AI video for TikTok
How distribution actually escalates (practitioner-aggregated, not TikTok's own published documentation): a new video is shown to a small cold-start pool first, measured over the first 30-90 minutes. Clearing roughly 35% completion plus at least one real engagement signal graduates it to a much wider pool; falling short means the video doesn't get a second look — there's no slow build on TikTok, only pass or stall. The bar to go genuinely viral has reportedly risen over time as the platform matured.
Register: first-person, in-the-moment, imperfect, sound-led. It dies when it looks like an ad from another platform.
Anti-pattern: broadcast polish; no sound design on the one platform where sound sells; a hard cut to black that kills the loop.
TikTok persuasion skeleton
0–2s: anomalous or lived open (sound-led) + brand/product already in world → 2–8s: one visible proof beat inside the same moment → last 2–3s: one spoken or on-screen CTA that matches the proof.
e.g. Open mid-use, not mid-slogan; the click lands after the result is visible, not before.
AI video for Instagram and Facebook
One underlying platform and ranking system (Meta's Andromeda) — placements differ in frame and hook window, not in fundamental behavior.
Register: crafted but casual on Reels; composed and still-image-logic on Feed; direct address on Stories; broad and plain on Facebook proper.
Common failures per placement
Symptom — Same platform, different traps depending on where it runs
- On Reels — reposting a TikTok with its watermark still visible, or putting the CTA in the bottom third where the platform UI covers it (sends, not likes, are what actually buy reach here). On Feed — uploading 16:9 into a 4:5 slot and surrendering roughly 40% of the visible area. On Stories — telling one story across multiple cards; every tap is an exit, so one card needs to carry one whole idea. On Facebook proper — subculture references for a broad, older, sound-off audience that didn't choose to be watching.
AI video for YouTube
Five distinct ad units in one ecosystem, bound together by the ABCD framework (Attention, Branding, Connection, Direction) and one delivery spec.
Register: earn the first 5 seconds honestly on in-stream — entertainment is the currency. It dies when it opens with a logo sting. Long-form wants documentary, tutorial, or genuine story — it dies when it's clearly a 4-minute ad.
Anti-patterns: building to a reveal at 0:22 on skippable in-stream (the reveal has to exist in some form before 0:05); attempting a narrative in a 6-second bumper (it's one image plus one brand moment, not a story); horizontal footage blurred-and-boxed into a vertical Shorts frame.
Skippable ABCD in practice
Attention (0–5s: anomalous or useful image, brand present) → Branding (name/product already readable) → Connection (proof demo or specific outcome) → Direction (one CTA that matches the proof).
e.g. If the skip hits at 0:05 and they still don't know the claim, the cut failed before Connection ever started.
AI video for LinkedIn
The sound-off, professional-register outlier — captions carry the hook, not audio.
Register: authoritative and human — most B2B buyers say thought leadership can be rigorous and engaging at once [research]. It dies when it's either a press release or a TikTok dance dropped into the wrong feed.
Anti-pattern: consumer-influencer register in a professional feed; no captions on a platform where most viewers never turn sound on; a four-minute talking head with no claim in the first three seconds.
LinkedIn persuasion skeleton
0–3s: captioned claim a peer would respect → mid: one specific proof beat (number, demo, or lived consequence) → close: one professional next step that matches the proof.
e.g. "We cut review cycles from five days to one" on screen, then the process that caused it — then the ask.
AI video for Pinterest
[trade]Register: planning, not scrolling — closer to search intent than entertainment. It dies when it's built for passive consumption instead of a decision someone's actively making.
AI video for Snapchat
Native, camera-first — and the one platform where a required loudness spec coexists with a sound-off-leaning viewer default.
Register: native, young, playful, camera-first. It dies when it looks polished.
Anti-pattern: anything that looks produced; wasting the first non-skippable seconds on a slow build.
AI video for X (Twitter) and Threads
Both text-first, conversational, sound-off feeds. Threads has no independently published video spec of its own — treat it as inheriting Meta's Feed behavior until one exists.
[trade].
Don't default to Threads for a video-first campaign — treat it as an add-on placement, not a primary
one.Register: X wants fast, opinionated, quotable, in-joke-fluent — it dies when it's earnest and slow. Threads wants conversational and text-first — it dies when it's a video-only ad with no caption carrying the weight.
AI video for WhatsApp Status, Twitch, and website heroes
Three small, distinct destinations, bundled because none carries enough of its own data to justify a separate page.
Website hero anti-pattern: a heavy, multi-megabyte 4K loop that becomes the page's largest paint element and slows the whole page down; a visible loop seam; an audio track left on a video that plays muted anyway; motion that competes with the headline instead of sitting behind it.
AI video for broadcast TV, CTV/OTT, and cinema
The no-skip, sound-ON, big-screen tier — sharing one exit-pressure profile (none) and a seriousness about QC and loudness compliance social platforms don't have.
India-specific compliance note: every new TV advertisement requires formal certification before broadcast, on top of standard advertising-code compliance — a real, dated regulatory requirement, not a stylistic recommendation. UK requires equivalent clearance-body approval before transmission. [official]
Anti-patterns: audio present on the clock/slate before an Indian broadcast spot (an automatic QC failure); missing loudness targets on US/UK delivery; mobile-scale on-screen type that's unreadable across a living room on CTV; a web-grade master upscaled for cinema, where compression artifacts and soft focus are twelve meters wide.
Built from official sources
- Platform-specific figures above carry inline confidence tags matching the internal source's own convention: `[official]` = the platform's or a standards body's own published documentation, `[research]` = a named, citable study, `[trade]` = practitioner/agency consensus with no controlled study located, `[contested]` = credible sources disagree and the disagreement is stated rather than resolved. Don't quote a `[trade]` figure to a reader as if it carried `[official]` weight.