/director

Hooks, Cliffhangers & Open Loops — The Engagement Engine

Story is the content of a film; attention is the substrate it runs on. You can have a flawless three-act structure and lose the entire audience in the first 800 milliseconds. This chapter is about the substrate: the cognitive mechanics of grabbing attention, holding it across time, and engineering the moments where a viewer decides — consciously or not — to keep watching. We separate what is rigorously true (information-gap theory, the social-feed swipe decision) from what is folklore dressed as science (the "Zeigarnik effect" as usually told), because building on a myth produces brittle work.

A note on scope: the sibling skill cinematic-ai-video already ships the arousal-arc templates and platform manipulation tables. This chapter is the foundation under those tables — the why, the honest science, and the shot-list-level execution. For how attention interacts with cutting rhythm see 10-editing-theory.md; for the larger narrative scaffolding hooks hang inside, see 01-story-structure.md.

The first-frame contract

Every shot makes a promise. The first frame contract is the implicit deal struck in the opening moment: this is the kind of thing you're about to watch, and here is the question that will be answered if you stay. The viewer evaluates the promise, then either commits or leaves. The job of the opening is not to be impressive — it is to make a promise interesting enough that the cost of leaving feels higher than the cost of staying.

This reframes a common beginner error. New filmmakers open with establishment (here is the world, here is the weather, here is a slow drone shot of a city). Establishment answers a question nobody asked yet. The contract is strongest when the first frame raises a question rather than settles one.

Contrast two openings of the same hypothetical scene — a man in an apartment:

The second frame withholds the establishing answer and substitutes a generative question. This is the entire mechanism of the cold open and in medias res ("into the middle of things," Horace's phrase from Ars Poetica): start after the inciting event so the audience must reconstruct the world while a question pulls them forward. Breaking Bad's pilot opens with pants floating in the desert air and an RV careening through the dust — we spend the rest of the act earning the explanation. Saving Private Ryan drops you on Omaha Beach before you know a single name.

The deep reason in-medias-res works is load order. Human attention allocates to unresolved states far more readily than to resolved ones. A settled world is cognitively "done" — nothing to track. An unsettled world opens a tracking slot in working memory, and an open slot is itself a low-grade form of arousal that the brain wants to close.

→ AI APPLICATION. Generative video models reward this because they are still weak at long, uneventful continuity but strong at a single charged moment. Do not prompt a model (Veo 3.1, Kling 3.0, Runway Gen-4.5 — see the model table later) for "a man wakes up and gets ready for work." Prompt for the charged fragment: extreme close-up, trembling hands scrubbing dried blood from under a fingernail over a steel sink, harsh overhead kitchen light, shallow depth of field, 35mm, tension. Then write the answer into later shots. In your shot list, tag the opening shot CONTRACT: and state in one line the question it must raise. If you can't write that line, the shot is establishment and should be cut or moved.

The swipe decision: 0.3–3 seconds

On a social feed the contract is evaluated brutally fast. Eye-tracking and platform retention data converge on a swipe decision made in roughly 0.3 to 3 seconds — closer to the low end on a fast scroller. This is not a metaphor; it is a measured behavior with hard algorithmic consequences.

The numbers, as of 2025 industry retention data (these are platform-reported / aggregator figures, treated below as directional, not laws of nature):

• Over 70% of users decide to keep watching or scroll within the first 3 seconds.
• Videos holding 70–85% retention through the first 3 seconds receive roughly 2.2× the total views of lower-retention videos.
• Strong hooks hold 80–90% past 3 seconds; weak hooks hold only 30–40%.
• Below ~60% 3-second retention, the algorithm gives minimal distribution — the video effectively dies regardless of likes.

The mechanism behind the multiplier is the feedback loop: platforms (TikTok, Reels, Shorts) use early retention as a quality proxy to decide whether to widen distribution. A weak first 3 seconds is not just "fewer viewers" — it is a signal that throttles every downstream impression. This is why the first 3 seconds matter disproportionately on social and far less in a theater, where the viewer has already paid and sat down. The medium changes the contract's deadline.

A crucial honesty caveat: the exact percentages above circulate widely and are repeated across marketing blogs with suspicious precision. Treat the direction as solid (early retention strongly predicts reach; the first seconds are decisive) and the specific cutoffs as approximate, platform-specific, and constantly changing. Do not hard-code "the 60% rule" into a creative brief as if it were physics.

→ AI APPLICATION. For vertical social, the AI pipeline must front-load the single most arresting frame. Generate the hook beat first and independently, then build the rest around it. Concretely: render 4–6 variant openers (different first frames) and A/B the 3-second retention rather than trusting your taste. Use a motion-forward first frame — generative models default to slow drift, which is death on a feed. Prompt explicitly for immediate motion or a visual anomaly in frame one: opening frame: a hand already mid-slam onto a table, objects jumping, fast rather than a person sitting at a table who then slams it. The second phrasing wastes the swipe window on setup.

Curiosity, done rigorously: Loewenstein's information-gap theory

The popular term is "curiosity gap." The rigorous basis is George Loewenstein's 1994 paper The Psychology of Curiosity: A Review and Reinterpretation (Psychological Bulletin, 116(1), 75–98). His information-gap theory is the most defensible model of why a well-built hook works, so it is worth getting exactly right.

Loewenstein's claim: curiosity is a response to a perceived gap between what one knows and what one wants to know. The gap is experienced as a form of deprivation — an aversive state, like a mild itch — and the drive to resolve it is the drive to scratch. Three properties of his model matter enormously for filmmaking:

1. Curiosity requires a reference point. You must know enough to perceive that something is missing. A total stranger to a domain feels no gap because they don't know what they don't know. This is why a hook must give the viewer a frame and then a hole in the frame — not pure mystery (which is just confusion) and not full information (which closes the gap).
2. Curiosity is an inverted-U in the size of the gap. A tiny gap is boring (barely worth resolving). A vast gap is overwhelming (feels unresolvable, so the brain disengages). Peak curiosity sits at a medium gap — close enough to feel resolvable, far enough to feel worth resolving. This is the single most actionable insight in the chapter.
3. Curiosity intensity scales with proximity to resolution. The closer you feel to the answer, the stronger the pull. This is why hooks that imply "the answer is seconds away" outperform open-ended mystery.

The clickbait "curiosity gap" headline ("You won't believe what happened next") is a degenerate, cynical exploitation of property 1 — it manufactures a frame-and-hole with zero substance. It works once and trains the audience to distrust you. The craft version uses a real gap with a real payoff: the difference between a headline and a hook.

→ AI APPLICATION. When an LLM drafts or audits a hook, it should explicitly model the gap. Prompt the model: "State the viewer's reference frame after shot 1 in one sentence, then state the single missing piece of information that frame creates. If you cannot state both, the hook has no gap." For the inverted-U, instruct it to rate gap size 1–5 and flag anything rated 1 (boring) or 5 (overwhelming), targeting 3. This converts a vague note ("make it more intriguing") into a measurable, fixable spec.

The Zeigarnik effect: the honest version

This is where most engagement writing goes wrong, so we go slowly. The Zeigarnik effect is the popular claim that people remember interrupted or incomplete tasks better than completed ones — and it's cited everywhere as the scientific justification for cliffhangers and open loops.

Here is the honest state of the evidence. The effect comes from Bluma Zeigarnik's 1927 study under Kurt Lewin, in which interrupted tasks were recalled about twice as often as completed ones. The problem: it replicates poorly. A 2025 meta-analysis (Interruption, recall and resumption, Humanities and Social Sciences Communications) found no reliable memory advantage for unfinished tasks. Excluding Zeigarnik's original 1927 data, the interrupted-to-completed recall ratio was essentially 0.99 — i.e., no effect. The supposed memory boost is, by current evidence, not robust.

But — and this is the nuance that careless writers miss — the same meta-analysis found that a different effect does hold up: the Ovsiankina effect, a real and general tendency to resume interrupted tasks. People don't necessarily remember the unfinished thing better, but they are genuinely driven to complete it. The drive is real; the memory claim is not.

So when someone tells you "cliffhangers work because of the Zeigarnik effect," the accurate correction is:

The memory story is overstated and largely failed replication. But the completion drive (Ovsiankina) is real, and it — combined with Loewenstein's information-gap deprivation — is the better-supported reason an open loop pulls a viewer forward.

The narrative technique still works. It just works for the correct reason: an unresolved question is an aversive open state the viewer is motivated to close, not because they remember it better but because the open loop generates a small, persistent pull toward resolution. Build on the mechanism that survives scrutiny, not the headline that didn't.

→ AI APPLICATION. Do not let an LLM cite "Zeigarnik effect" as load-bearing justification in a creative rationale — it will parrot the pop-psych version. Add a guard to the system prompt: "If invoking the Zeigarnik effect, note that the memory claim failed to replicate (2025 meta-analysis) and that the durable mechanism is the resumption drive (Ovsiankina) plus Loewenstein's information gap." This keeps the bible's honesty standard intact and prevents you from designing around a false premise — e.g. don't assume the audience will remember a dangling thread across a 6-shot gap; they may not, so re-surface it rather than relying on recall.

Open loops and the re-hook cadence

An open loop is a deliberately unresolved question kept alive in the viewer's mind. The first-frame contract opens the master loop; sub-loops open and close throughout. The art is in cadence: you want loops to overlap so that one is always pulling forward, never letting the viewer reach a fully resolved state until you choose.

On long-form (theater, episodic), loops can run for an hour. On social, attention leaks fast and must be re-baited continuously. The widely-cited social practice is a re-hook every 5–8 seconds — a new visual, a turn, a fresh micro-question — to reset the swipe decision before it can fire. The 5–8 second figure is heuristic, not measured law, but the principle is sound and follows directly from the swipe mechanics: if the viewer never reaches a "nothing-new-is-coming" state, they never re-evaluate leaving.

A useful frame is the retention curve: plot % of viewers still watching against time. Two features predict performance — the initial drop (slope in the first 3 seconds) and mid-video leaks (sudden cliffs where many viewers leave at once). A leak is almost always a closed loop with no new loop opened — a moment where the viewer's question got answered and nothing replaced it. The fix is structural, not cosmetic: open the next loop before you close the current one.

→ AI APPLICATION. Build the re-hook cadence directly into the shot list as a column. For each 5–8 second block, write the new question or visual turn that resets attention. When auditing, ask the LLM: "Walk the script in 5-second windows. For each window, name the open question driving the viewer forward. Flag any window with no open question — that is a predicted retention leak." This turns the abstract retention curve into a concrete per-window checklist before a single frame is generated.

Cliffhanger taxonomy

"Cliffhanger" is not one device. Naming the variants lets you choose deliberately. The term itself comes from serialized Victorian fiction (Thomas Hardy's A Pair of Blue Eyes literally left a character hanging from a cliff).

The streaming platforms turned this taxonomy into infrastructure. Netflix's post-play / auto-play is the clearest case of engineering wrapped around craft. When an episode ends, a countdown (rendered as a color wipe across the "Next Episode" button) auto-starts the next episode in about 5 seconds unless the viewer intervenes. The genius — and the manipulation — is that it converts the cliffhanger's pull into a default: continuing requires no action, while stopping requires an active choice. Netflix has stated auto-play produced the single biggest increase in hours-watched of any feature it tested. Episode endings are now written to this mechanic: end on an open loop precisely because the platform will start resolving it before the viewer can decide to leave. (Note that the platform later added the ability to disable auto-play, after well-founded criticism that the default exploited the absence of a stopping cue — an honest acknowledgment that this is closer to a dark pattern than pure craft.)

→ AI APPLICATION. For AI-generated series or multi-part social content, design the last beat of each part as a typed cliffhanger and the first beat of the next part as the matching resolution-tease. In the shot list, tag the final shot with one of the five types above and write its open question. If you're publishing to a feed, exploit the platform's own auto-advance the way Netflix does: end part N on a dangling threat whose answer is the literal first frame of part N+1, so the algorithm's auto-scroll does the "post-play" work for you.

Pattern interrupts and the causality test

Two more tools complete the engine.

A pattern interrupt is a deliberate break in the established rhythm — a sudden cut, a sound drop, a tonal flip, an unexpected angle — that resets habituating attention. The brain habituates fast: a steady visual rhythm becomes predictable, prediction lowers arousal, and lowered arousal leaks viewers. An interrupt forces the prediction machinery to re-engage. Edgar Wright's whip-pans and hyper-cuts in Hot Fuzz and Baby Driver are pattern interrupts as a stylistic engine. Used too often, they become the new pattern and lose their power — interrupt scarcity is what gives them force.

The causality test (Trey Parker and Matt Stone's "but / therefore" rule, given to an NYU class in 2011) is the most practical structural filter for engagement. Walk through your beats. If the connective tissue between two beats is "and then," the structure is dead — you have a sequence of events with no momentum. Every beat should connect with "therefore" (logical consequence) or "but" (complication / reversal). Parker's exact framing: if "and then" sits between your beats, "you're fucked."

The reason this is an engagement tool, not just a structure tool, is that causal links are themselves open loops. "Therefore" carries the question what is the consequence?; "but" carries how is this resolved? "And then" carries no question — it's just the next thing — so attention has nothing to ride. Causality and curiosity are the same force seen from two angles.

Bad:  He finds the letter. AND THEN he goes to work. AND THEN he calls his sister.
Good: He finds the letter, BUT it's addressed to a dead man, THEREFORE he tracks
      the sender, BUT the address is his own house.

Each "but/therefore" plants and partially pays a loop; the "and then" version is a flat retention curve waiting to leak.

→ AI APPLICATION. This is the single best mechanical pressure-test to hand an LLM. Prompt: "Rewrite the beat sheet inserting the literal word 'but', 'therefore', or 'and then' between every pair of beats. List every 'and then' — each is a causality break and a predicted attention leak. Propose a 'but' or 'therefore' replacement for each." Combine with the 5-second-window audit above and you have a two-pass leak detector: one pass for causality (structural pull), one for re-hook cadence (moment-to-moment pull).

Putting it in a shot list

The deliverable of this chapter is a shot list whose first column is attention, not just picture. Add three fields to your normal shot list:

Then run the LLM audit as the final gate before generation:

1. First-frame check — does shot 1 raise a question instead of settling one? Rate the information gap 1–5; target 3.
2. Window walk — in 5-second windows, name the open LOOP; flag empty windows.
3. Causality scan — find every "and then"; demand a but/therefore.
4. Cliffhanger tag — each part/episode ends on a typed cliffhanger with a stated open question, and the next part opens on its tease.
5. Honesty guard — no rationale leans on the (debunked) Zeigarnik memory claim; pull is justified by information-gap deprivation + resumption drive.

The current model landscape for executing the resulting beats (verify — fast-changing; figures as of mid-2026 web research):

The craft principle is constant across whatever model wins next year: open a real gap, keep a loop always pulling, link every beat with but/therefore, and never trust a swipe-window to forgive a slow start.

Sources

• Loewenstein, G. (1994). The Psychology of Curiosity: A Review and Reinterpretation. Psychological Bulletin 116(1), 75–98. https://www.cmu.edu/dietrich/sds/docs/loewenstein/PsychofCuriosity.pdf
• Interruption, recall and resumption: a meta-analysis of the Zeigarnik and Ovsiankina effects (2025), Humanities and Social Sciences Communications. https://www.nature.com/articles/s41599-025-05000-w
• TikTok first-3-seconds retention statistics. https://insights.ttsvibes.com/tiktok-first-3-seconds-hook-retention-rate/
• Hootsuite, How the TikTok algorithm works in 2026. https://blog.hootsuite.com/tiktok-algorithm/
• Binge Logic: How Streaming Platforms Engineer Cliffhangers (2026), MediaMikes. https://mediamikes.com/2026/03/binge-logic-how-streaming-platforms-engineer-cliffhangers-for-maximum-retention/
• HN thread from the dev who built Netflix autoplay. https://news.ycombinator.com/item?id=20566514
• David Perell, The But & Therefore Rule. https://perell.com/note/but-therefore-rule/
• Best AI Video Generators 2026: Sora 2 vs Veo 3.1 vs Kling 3.0 vs Runway. https://www.getaiperks.com/en/blogs/44-best-ai-video-generators-2026
• After Sora: Best AI Video Generators 2026. https://www.digitalapplied.com/blog/after-sora-best-ai-video-generators-2026-runway-kling-veo