Cliffhanger Engineering for Serialised Verticals

A cliffhanger in a 60-second vertical isn't doing the same job it does in a television drama. TV can afford a slow build to the cut because the next episode is already scheduled — the audience isn't going anywhere. A vertical series has no such guarantee: the very next swipe is a different creator's video, and the only thing standing between "watched episode one" and "never saw episode two" is whether the cut itself was strong enough to make someone stop scrolling and tap through. That's not a writing flourish. It's the entire retention mechanism the format runs on, and it lives or dies on one timing decision: exactly where the cut happens relative to the beat it's interrupting.

The cut has to land before the beat resolves, not after

The instinct when building a cliffhanger is to build toward it — set up the tension, let it climb, and cut at the peak. That's usually one beat too late. Versely's own retention-curve mechanics describe the failure pattern precisely: "a flat section followed by a drop points at one specific beat — usually the moment the video stops delivering and starts explaining." A cliffhanger that lets the tension peak and then lingers for even a beat while the scene explains what just happened has already crossed into the moment viewers start leaving. The cut needs to land while the curve is still rising — mid-reveal, not post-reveal; the instant before the danger, not the instant it's confirmed.

Practically, that means cutting earlier than feels comfortable in the edit. If the natural instinct is to let a line of dialogue finish before the cut, try cutting mid-sentence instead. If the instinct is to show the door opening and then cut, try cutting on the door starting to open. The discomfort of an "unfinished" cut is exactly the mechanism — an audience that got a complete beat has no reason to come back for the rest of it, and an audience that got an interrupted one does.

Cliff types that hold up at 60 seconds

Not every kind of tension survives a runtime this short. A slow-burn mystery cliffhanger that needs two minutes of setup to land doesn't fit in a format where the whole episode might be sixty seconds total. Five cliff shapes reliably compress into that runtime:

• Interrupted reveal. A face turns toward camera, a document starts to open, a text message starts to load — cut before the payload is visible. The audience has seen enough to know something was about to be revealed and not enough to know what.
• Posed question, no answer. A character asks a direct question mid-scene and the cut lands on the question, not the reply. This is the cheapest cliff type to write and one of the most reliable, because the open loop is explicit rather than implied.
• Imminent danger, no outcome. A hand reaches for a door handle, a car starts to swerve, a phone starts to ring at the worst possible moment — cut on the action starting, not the consequence landing.
• Identity interrupted. A character starts to say a name, or a masked figure starts to turn around — cut on the motion, before the identity confirms. This works especially well as a season-level hook, not just an episode-level one.
• Arrival interrupted. Someone is about to walk into a room where they're not expected, a car is about to pull into a driveway — cut on the approach, not the entrance.

All five share the same shape: they cut on the setup of a beat, never the payoff. The moment the payoff would land is the moment episode two exists to deliver, and that's the actual product being sold to the viewer who taps through.

Building the mechanical cut

The technical version of "the cut has to land before the beat resolves" is a job for first-last-frame generation — a method that takes a start image and an end image and generates the motion connecting them, rather than improvising an open-ended clip. For a cliffhanger specifically, that mechanism is useful in two different ways.

The first is building the cliffhanger shot itself: set the start frame at the beginning of the interrupted beat — the hand approaching the handle, the face starting to turn — and set the end frame at the exact interrupted moment you want the episode to freeze on, rather than letting a generation run past that point and trimming it in the edit. This is more reliable than trimming after the fact, because the model is generating motion that's built to land exactly where you specify, rather than motion aimed at a natural conclusion that you then have to cut off mid-flow.

The second use is the transition between episodes. Black Forest Labs describes FLUX 3 Video's approach to this directly: you can start from an image, specify an end frame, or set multiple keyframes in a clip, and the model connects them in sequence. Setting the last frame of episode one's cliffhanger as the first frame of episode two's cold open produces a transition that reads as continuous rather than as two separately generated clips stitched together — the visual equivalent of the cliff not actually resolving until the viewer is already inside the next episode.

The one failure mode worth watching for either use: the same glossary entry warns that the two frames have to be plausibly connectable in the available time — a start and end frame from wildly different compositions won't produce a camera move, it'll produce a warp, and a warp reads as an editing error rather than as tension.

Which model to reach for

Versely's catalog carries three first-last-frame models worth knowing for this specifically: Flux 3 First Last Frame to Video, VEO First Last Frame, and VEO First Last Frame Fast. All three generate native audio alongside the picture, which matters for a cliffhanger built around a sound cue — a phone starting to ring, a door starting to creak — since the audio and the interrupted motion are generated together rather than synced up afterward. VEO First Last Frame Fast is the lower-cost, quicker-turnaround option worth using while you're iterating on exactly where the interruption point should sit; the standard VEO and Flux 3 versions are worth reserving for the final pass once the timing is locked.

Walkthrough: engineering one cliffhanger

1. Pick one of the five cliff shapes above and write the beat it's interrupting — the full version, payoff included, even though the payoff won't be shown.
2. Identify the exact frame where the interruption should land — not the natural end of the action, but the point mid-action where continuing would start to resolve it.
3. Generate the start frame (the beginning of the beat) and the end frame (the interruption point) as two stills first, before generating any motion — this is the fastest way to confirm you've picked an end frame that's genuinely mid-action rather than accidentally a resolution.
4. Ask Versely's agent to build the transition between the two — attach both stills and it picks a first-last-frame-capable model and generates the interpolated motion between them.
5. Watch the result at the same speed a viewer would — full playback, once, no scrubbing — and check the honest question: does it stop clearly before the payoff, or did it sneak past into resolution? If it resolved, trim the end frame earlier and regenerate rather than cutting the finished clip; a cut clip still shows the motion completing right up to where you trimmed it, which reads differently than motion that was generated to stop exactly there.
6. Set the same end frame as episode two's opening first frame if you're chaining a full series, so the cut across episodes reads as one continuous interruption rather than two separate pieces of footage.

The whole discipline here is resisting the edit's natural pull toward completion. Every instinct in cutting a scene points toward finishing the beat; a cliffhanger only works if you fight that instinct on purpose, and the models above are the tool for making the interrupted version look intentional instead of unfinished. Versely's full first-last-frame and image-to-video roster covers every other angle on the same mechanic, cliffhangers included.