How to script a Short that holds attention

What the feeds measure

Most Shorts are watched in the Shorts feed, where there is no thumbnail, so the opening seconds do the job a title and thumbnail do on a long video. YouTube describes a Short's first views as working like impressions, with the swipe in place of the click (YouTube Official Blog), and its analytics report "stayed to watch", the share of times viewers stayed past the first seconds of a Short (YouTube Help). TikTok says whether someone finishes a longer video is a strong signal that outweighs weak ones. Instagram says it predicts whether you will watch a reel all the way through, reshare it, like it and visit its audio page.

The first seconds predict the rest. In a study of 90,000 short videos from Snapchat, the share of viewers who watched at least 5 seconds correlated 0.926 with how much of each video people watched on average (Li et al., ECCV 2024). That makes the opening 5 seconds the best single place to spend effort.

None of them reward length on its own. A longer Short only helps when people keep watching it, so the script's job is to earn each next second. That comes down to three places: the first line, the middle and the end.

The first line names the thing

Curiosity is a response to a specific gap in what someone knows (Loewenstein, 1994). A vague teaser leaves nothing specific missing, so it creates no gap. The first line names the object, the number or the problem, and the rest of the video fills in the answer. YouTube's own advice for Shorts is to "start with a question, a surprising fact, or a visually captivating moment" and to "get to the point quickly" (YouTube Official Blog).

• Weak: "A lot of people ask me about saving money, so let's talk about it."
• Strong: "Your paycheck should land in the account you never touch."

• Keep it to 6 to 14 words and one claim.
• Get the topic noun or a number into the first six words.
• For advice, frame it around the viewer (you, your) and a problem they already have.
• Start on the first word: no greeting, no name, no channel intro, no breath before speaking.
• Put the idea on screen as a title of 3 to 8 key words, not a copy of the spoken line, for people watching with the sound off.

The on-screen line matters because many people watch muted. In a 2019 survey of 5,616 US adults, 69% said they watch video with the sound off in public (Verizon Media and Publicis Media, reported by Streaming Media). Keep the title to key words: a review of 63 studies found that narration with short written key words helps people learn, while narration with the same words written out in full gets in the way (Trypke, Stebner and Wirth, 2023).

The second line says what the end holds

By about the fifth second, tell the viewer what they will get and how far along they are. A count works ("three accounts and one rule"), and so do a rating, a race against a clock, a before and after, or a reveal you have shown blurred. TikTok's advertising guidance makes the same point in its own terms: introduce the proposition in the first 3 seconds and land the hook within the first 6 (TikTok, Creative best practices). Meta calls this a statement of intent hook, "telling viewers exactly what they're about to see and learn", alongside value promises and questions (Meta for Business).

Then pay something off early. By about the eighth second the viewer should have a first fact, number or step, not only setup.

Word budgets

Write to a target length. Jump cuts remove pauses and breaths, so an edited talking head plays faster than the same words read aloud. These budgets are for the finished cut:

When a draft runs long, cut a beat rather than squeezing more words into the same seconds. Captions have to stay readable: silent reading of non-fiction averages about 238 words a minute (Brysbaert, 2019), roughly 4 words a second, and text on a Short competes with speech and motion.

The middle: every beat adds something

• Two to four beats of 3 to 8 seconds each. Each one adds a new fact, step or turn, and no line repeats the hook or an earlier line.
• Join beats with but, so, therefore or which means. A chain of "and then" reads as a list with no tension.
• Order is a judgment call. We put the second-best point first and the best one last, with one line before it that raises the stakes and still carries a fact ("the third one is what stops the overspending"). In our own blind comparison of 8 scripts, a bare tease such as "this one is sneaky" read as filler and lost.
• Script the picture as well as the words. A beat lands harder when the screen adds something the voice does not say: the document, the number, the object.
• Change something on screen every few seconds, tied to the words: a graphic, a cut-in, a label.
• In a longer Short, open a fresh question about every 15 seconds. A complication before the answer (why the obvious fix fails, or the catch) keeps the answer uncertain until it lands.

The ending: deliver, then stop

Deliver exactly what the first line promised, before the last 15% of the runtime. Say at most one more line, then end. Silence after the last word gives the viewer nothing to stay for, so the last frame should come a fraction of a second after the final syllable.

The end is the second place people leave. In a study of TikTok-style viewing with 133 online participants and 15,344 swipes, 29% of swipes came in the first 20% of a video and 42% in the last 20%, a count that includes moving on once a video had ended (Li et al., NSDI 2023). Skip the wrap-up line ("so remember", "the bottom line"): say the last useful thing and stop.

A loop is optional, only where it reads naturally: an ending that leads back into the first line, or the shot you opened with. In our own tests, loop and callback endings did not beat a clean stop, so ending right after the payoff stays the rule. Never hold back the payoff to fake a loop. On YouTube, replays count as views but not as engaged views, so a loop lifts the public count more than the number YouTube uses for the Partner Program (YouTube Help).

Calls to action

• One at most, after the payoff, in one sentence. Or none at all.
• Match it to the goal: a send for reach that names the person ("send this to whoever pays the rent", never just "a friend"), a real either-or question for comments, a save for a tutorial, a follow only when a series exists.
• Skip engagement bait. Facebook demotes posts that goad people into likes, shares or comments, and since January 2019 that includes bait spoken in a video's audio (Meta). Asking for help, advice or recommendations is fine.

A 30-second example

The count in the second line is the promise made visible: the viewer can see how close the end is. The close returns to the opening shot because it reads naturally here, not because a loop is required.