How to use
- Copy the installation command and review the source package before installing it in your own agent environment.
- Check the source document for required tools, credentials and supported hosts.
- Describe your input and desired deliverable; review the output against the source material.
npx skills add https://github.com/apimageorg/apimage-skills --skill video-hook-first-3-secondsBefore you start
- Installation does not include a model subscription, API allowance or permission to access your files.
- Repository stars and installation counts describe activity, not an independent quality or security test.
Rendered from the source repository. Headings, examples, code, tables, links, and referenced images are preserved.
The First Three Seconds
On every short-form platform, the opening frames decide the outcome. Retention curves fall off a cliff in the first two to three seconds and the algorithm reads that fall-off as a verdict on the whole video.
Which means the hook isn't the intro to the video. The hook is the video's audition, and everything else only gets watched if it passes.
What the first frame has to do
Three jobs, simultaneously, before anyone has decided to watch:
- Stop the scroll — visual interruption, unexpected motion, or a face
- Establish stakes — why this matters to me, right now
- Promise a payoff — something is going to resolve, and I want to see it
A clip that opens on a slow establishing shot of a product on a white background does none of the three. It reads as an advert in about 400 milliseconds, and the thumb moves.
Hook patterns that survive
| Pattern | Opening frame | Works because |
|---|---|---|
| Mid-action start | Already in motion, no setup | No dead frames. Motion holds the eye |
| Result first | The finished outcome, then "here's how" | Payoff promised immediately |
| Problem in frame | The visible mess, break or failure | Recognition. That's my problem |
| Pattern interrupt | Something visually wrong or unexpected | Curiosity is involuntary |
| Direct address | Face, eye contact, talking | Faces are pre-attentive. Hard to scroll past |
| Extreme close-up | Texture, detail, unclear at first | Ambiguity resolves at 2s, buying you the watch |
| Countdown or list | "Three things about..." | Structure implies a finite, worth-it commitment |
Mid-action is the most reliable and the most under-used. Most generated video opens on a static or slowly-moving establishing shot because that's what the prompt implied. Explicitly prompting for motion that's already underway at frame one removes the dead opening.
Prompting for it
The default output of a text-to-video prompt tends to start calm and build. You have to ask for the opposite.
Weak — starts static:
A skincare bottle on a marble counter, camera slowly pushes in,
soft morning lightStrong — starts mid-motion:
Hands already mid-pour, cream falling from bottle to palm, motion
blur, close-up, camera locked. Action is underway at the first frame,
no lead-inThe phrases that do the work: already mid-action, starts in motion, no lead-in, action underway at frame one, cut in on movement. State them explicitly — the model will otherwise give you the establishing shot.
The generation call
generate_video(
mode="image-to-video",
model="flux-3-video-draft",
reference_images=["<product or scene still>"],
prompt="Hands already mid-motion at frame one, no lead-in. Fast "
"close-up, subject fills frame, natural motion blur. Camera "
"locked, no drift. Energetic, handheld feel.",
aspect_ratio="9:16",
resolution="hd",
duration=4,
seed=1234
)Generate the hook as its own 4-second clip. Not as the opening of a 15-second render. Reasons:
- Four seconds is the cheapest unit to iterate on
- You can test six hooks against one body for the price of two full clips
- Hooks are the highest-variance element, so they deserve the most iterations
- Assembling hook plus body is a cut, not a re-render
See multi-scene-video-assembly and video-ab-testing-variants.
Test the hook in isolation
The hook is judged in isolation by the viewer, so judge it that way yourself.
Watch the first three seconds muted, on a phone, at arm's length, in a feed if you can simulate one. That's the real viewing condition. A hook that only works at full attention on a desktop monitor with sound isn't a hook.
Three checks:
- Muted test. Most feed viewing starts silent. Does the opening make sense with no audio?
- Thumbnail test. Freeze frame one. Would you stop on that image alone?
- Three-second cut. Show only the first three seconds. Does someone ask what happens next?
If the answer to the third is no, the rest of the video doesn't matter yet.
Common failures in generated video
- The slow push-in opening. The default, and the weakest. Prompt against it explicitly.
- Text-on-screen doing all the work. If the visual is inert and the hook is a caption, the visual is failing.
- A logo in frame one. Reads as an ad instantly. Brand later, once attention is earned.
- Empty frames. A composition with the subject small and lots of background has nothing to grab.
- Motion that starts at second two. The model rendered a lead-in you didn't ask it not to.
- Beautiful but static. Aesthetics don't stop scrolls. Motion and stakes do.
Iterate on the hook, not the video
Once you have a body clip that works, treat the hook as the variable. Six hooks, one body, same seed on the body — that's six testable creatives for a fraction of six full renders.
The hook is where almost all the performance variance in short-form lives. Spend the iteration budget there.
Don't
- Don't open on an establishing shot. Start mid-action.
- Don't put the logo in the first second.
- Don't judge the hook with sound on at full attention. Test it muted, on a phone.
- Don't render the hook inside a long clip while you're still iterating on it.
- Don't rely on a caption to carry a static visual.
- Don't iterate the whole video when the hook is the thing that isn't working.
Setup
claude mcp add --transport http apimage https://mcp.apimage.org/mcp \
--header "Authorization: Bearer sk_your_api_key" -s userRelated skills
tiktok-video-generation, multi-scene-video-assembly, video-ab-testing-variants, ugc-hook-library