Watch a retention curve for any short-form video ad and the shape is always the same: a cliff, then a slope. The cliff happens in the first one to three seconds and takes most of the audience with it. Whatever happens after that - the demo, the offer, the call to action - only ever happens to the people who survived the cliff. A 20% improvement in three-second retention lifts every downstream number at once, and no amount of polish applied to second fifteen does anything comparable.
This is well known and still routinely ignored, because the hook is one sentence and the rest of the script is twenty. Effort tends to distribute by word count rather than by leverage. The fix is structural: treat the hook as a separate artefact with its own testing loop, and write it against a taxonomy instead of writing whatever the brief happens to open with.
Test twenty hooks, not one
siriusly.ai generates hook variants as separate short renders, so you can find the opening that holds attention before you pay to render the full reel.
Try it freeWhat a Hook Actually Has to Do
A hook is not an attention grabber in the loud sense. Feeds are full of loud, and loud is cheap to produce, so viewers have long since learned to scroll past it. What a hook has to do is much more specific: give the viewer a reason to believe the next ten seconds are about them.
That breaks down into three jobs, all of which have to happen before the second sentence:
| Job | The viewer's question | How the hook answers it |
|---|---|---|
| Recognition | Is this about me? | Names a situation the viewer is in, in their words |
| Tension | Why keep watching? | Opens a gap - a claim unproven, a result unexplained |
| Permission | Is this worth trusting? | Sounds like a person talking, not a brand announcing |
Most weak hooks fail on recognition. They open with the product, the brand or the offer - all of which are answers to a question the viewer has not asked yet. A hook that opens on the viewer's situation earns the right to mention the product at second six.
Six Hook Mechanisms
Mechanism is how the hook opens, independent of what it says. It is the axis with the highest ratio of impact to production cost, because the same script body can be attached to any of them. Six cover most of the usable space:
1. The situation cold open
Start mid-scene, as if the camera caught the middle of something. "So I have been doing this every morning for two weeks and I need to tell someone." No setup, no greeting. The missing beginning is itself the tension.
2. The contradiction
State something that appears to argue against your own interest, then resolve it. "I do not think most people need this." Viewers are primed to distrust advertising, and a hook that concedes something disarms that reflex faster than any claim can.
3. The visual reveal
Lead with the thing being shown rather than said - a before state, an unusual object, a result held up to camera. Works best where the product's effect is visible, and fails badly where it is not, because a reveal that reveals nothing feels like a trick.
4. The direct address
Name the audience explicitly. "If you have ever cancelled a subscription and been charged anyway, this is for you." Blunt, and it works because it converts a broad impression into a self-selected one. The cost is reach: everyone outside the description leaves immediately, which is often exactly what you want.
5. The question that is not rhetorical
A real question with a real answer coming. "How much do you think this cost?" The rhetorical version - "Tired of feeling tired?" - is the most common bad hook in existence, because it has no answer and therefore opens no gap.
6. The correction
Overturn something the viewer probably believes. "Everyone told me to do X. It made it worse." Requires a belief that is genuinely common in your category, otherwise you are correcting no one.
Same script, three mechanisms
Cold open: "Okay, day fourteen, and I am genuinely annoyed that this worked."
Contradiction: "I was pretty sure this was a scam, so I bought it to prove it."
Correction: "Every single person told me to buy the expensive one. That was the mistake."
The body of the script - what the product does, the proof, the offer - can be identical in all three. Only the door changes.
Sentence-Level Rules
Once the mechanism is chosen, the writing itself follows a few rules that hold up unusually consistently:
- Under twelve words. Three seconds of natural speech is roughly ten to fourteen words. If the first sentence does not finish in that window, the hook is whatever fragment was audible, which is not what you wrote.
- The first three words carry the recognition. Captions are read faster than speech is heard, and the eye lands on the beginning of the line. Front-load the specific noun.
- No brand name in sentence one. Almost without exception. The brand is not the reason to keep watching unless the viewer already loves it.
- Concrete over abstract. "Forty minutes to fall asleep" beats "trouble sleeping". Numbers, objects and times are recognisable; categories are not.
- Write it as spoken, not as written. Read the hook out loud. If you would not say it to a friend, a generated creator saying it will sound exactly as staged as it reads.
Writing Hooks a Model Can Render
A generated hook has one extra constraint a filmed one does not: the model has to be able to perform it. This is where a lot of otherwise good hooks fall apart, and it is worth knowing the failure modes before you write twenty of them.
Hooks that fight the model: rapid physical action in the first second, a hard cut mid-sentence, multiple people talking, precise object interaction ("she twists the cap off and pours exactly one scoop"), or an emotional turn that has to land within half a second.
Hooks that render clean: a person mid-thought, a single held object, one clear expression change, a step toward or away from camera, and speech that starts on frame one instead of after a beat of silence.
The reason is worth understanding rather than memorising. Video models allocate their coherence budget to whatever is stable in the frame. A hook that demands three changes in the first second spends that budget on motion and pays for it with a face that drifts, which is precisely the second where the viewer is deciding whether this is a real person. Two of the failure modes above - drift and the silent lead-in - are the same ones that make chained generations look stitched, and the same fixes apply; the mechanics are in the last-frame chaining guide.
One more practical detail: leave no silence at the start. A model given a beat before the first word will fill it with a blink, a settle or a small head movement, and that beat lands exactly in the window where the viewer is deciding. Start the audio on the first frame.
Testing Twenty Hooks Without Burning the Budget
The whole argument for treating hooks separately is that they can be tested separately, and far more cheaply than full creatives. The workflow:
- Write the body once. One script, one persona, one offer. This is the constant.
- Write twenty hooks against the taxonomy. Not twenty rewordings - aim for at least three or four mechanisms represented, with several sentence patterns inside each.
- Render hooks only, at low cost. Three to five seconds each, at the cheaper quality tier. Twenty five-second clips cost a fraction of one full reel at maximum quality - the per-second numbers are in the model comparison.
- Judge on three-second retention, not on conversions. The clips do not contain an offer, so conversions are not the question. You are measuring one thing: did people stay.
- Attach the survivors to the full script and render properly. Usually two or three of the twenty. Those get the full-quality render, the captions and the localized versions.
The reason this ordering matters is that hook performance and body performance are close to independent. Discovering that a hook holds attention is cheap, and discovering it before you have rendered and captioned and localized twenty full reels saves most of the production budget for the creatives that deserve it.
How Hooks Fit the Wider Matrix
Hook mechanism is one of four axes worth varying in a creative matrix - alongside persona, awareness stage and creator. It is the cheapest of the four to vary and the fastest to read a result from, which makes it a good first axis, but it is not a substitute for the others. Twenty hooks on one persona still only ever reaches one kind of person. The full argument for why supply and variety became the dominant lever is in the piece on Meta's retrieval engine.
A useful sequencing rule: vary hooks within a persona to find the door, vary personas to find the room. Do the second only after the first has told you the room is worth entering.
Common Mistakes
Opening with the brand or the product name. The viewer has no reason to care about either yet, and the sentence spends its most valuable three words proving the video is an ad.
Fix: open on the viewer's situation. The product can arrive at second six, once there is a reason to hear about it.
Rhetorical questions. "Struggling with X?" opens no gap, because there is no answer coming and the viewer knows it.
Fix: ask a question with a real answer that arrives in the next sentence, or make a statement specific enough to be surprising.
Twenty hooks that are one hook. Rewording the same claim twenty ways tests synonyms, not mechanisms, and produces twenty statistically identical results.
Fix: require at least three mechanisms in any hook test. If two hooks would be answered by the viewer in the same way, they are one hook.
Judging hook tests on conversion rate. A five-second clip with no offer converts nobody, and a full creative's conversion rate mixes hook performance with everything downstream of it.
Fix: judge hooks on early retention alone, then judge the assembled creative on conversion. Two decisions, two metrics.
FAQs
How long should a hook be?
One sentence, under twelve words, delivered inside three seconds. If it needs two sentences, the first one is the hook and the second is already the body - which is fine, but write and test it that way.
Do captions change what the hook should say?
Yes, in one specific way: a large share of viewers read before they hear, so the hook is being consumed as text first. Front-load the concrete noun, and keep the line short enough that it does not wrap into three lines of burned-in caption. The mechanics of getting captions on screen word by word are in the ffmpeg and ASS guide.
Should the hook match the landing page?
The promise should, the wording does not have to. A hook that opens a gap the landing page never closes produces clicks and no conversions, which reads as a creative win and a business loss.
Does the same hook work across platforms?
Mechanisms travel; pacing does not. The same hook usually needs to be a beat faster on a feed where autoplay starts muted and the thumb is already moving. Retest rather than assuming, but do not rewrite from scratch.
How many hooks is too many?
The limit is statistical, not creative. Each hook needs enough impressions to separate it from noise, so twenty hooks on a budget that can only give each one a few hundred views produces twenty inconclusive results. Test as many as your traffic can resolve, and no more.