TikTok Hook Generator: Writing the First 3 Seconds That Stop the Scroll

The last video I posted got 4,200 views in 48 hours. The one before that: 1.1 million. Same account, same lighting, same script structure — the only meaningful difference was the first 1.7 seconds. I had spent about twenty seconds writing the second one, and maybe six minutes writing the first.

That gap is the entire job.

Most marketers treat the "TikTok hook" as inspiration. They open a doc, write "hey guys," stare at the ceiling, and ship whatever shows up. Then they blame the algorithm when the video lands at 800 views. The hook is not where you wait for the muse. It's a mechanical part of the work — and once you treat it that way, you stop being surprised by which videos win.

Here's the system I now use on every short-form script I write, and the AI workflow I run to draft 20 hooks in 15 minutes.

The first 3 seconds has three jobs, not one

Stop calling the first 3 seconds a "hook." It's three separate jobs running in parallel, and the best hooks do all of them at once:

1. Survive the thumb. You have about half a second before the user's thumb has already flicked. Your job is to break the automatic scroll, usually with a visual or a sound that doesn't match the rhythm of the previous video.
2. Signal genre. In the next second, the viewer is asking, "What kind of video is this — a story, a tutorial, a rant, a confession?" If you don't answer that in the first second, they don't care how good the rest is. They'll leave anyway.
3. Open a loop. In the third second, the viewer is unconsciously scanning for a question the video might close. If you don't open one, there's no reason not to scroll.

Most "good hooks" do job #1 with a visual gimmick. The hooks that consistently beat their channel's average also do jobs #2 and #3 in the same breath. The "I made $0 from TikTok for 60 days" video is doing all three: pattern interrupt (the dollar sign), genre signal (it's a story), and a loop (how did she get to $0, and what changed?).

If your hook only does one job, you're paying attention tax on jobs two and three. That's the difference between 4,200 views and 1.1 million.

Five patterns I actually use (and one I avoid)

I have a Notion page called "hooks" with 287 entries. 261 of them are bad, and I keep them to remind myself what bad looks like. Out of the 26 that survived, they cluster into five patterns. I'm going to name them and give the shape — not so you copy the lines, but so you have a stable vocabulary when you're writing your own.

The specific number. "I posted 47 TikToks in 30 days. Three of them broke 100k." Numbers are sticky because they make a claim falsifiable. A round number ("3 tips") reads as filler; an odd or specific number reads as a real person reporting from a real campaign. The pattern works best when the number is the result, not the count of items in the video.

The confession. "I have a $400/month ChatGPT habit and I can't justify it to my accountant." Confessions do the genre-signal job in one line: this is a personal video, not a brand video. They also pre-load the viewer with the assumption that you're going to be honest, which buys you trust for the rest of the script. They die when they're not actually confessing anything — "honest take" with no edge is just a sponsored post in a hoodie.

The pattern interrupt. Silence, a weird first frame, or opening on the payoff instead of the setup. "Here's the thumbnail that got me 2 million views" — said while showing the thumbnail, before any context. The pattern works because it skips job #2 (genre signal) entirely; the visual is the genre. The trap is that pattern interrupts age fast. Anything that felt fresh in 2023 now feels like a 2023 TikTok.

The call-out. "If you're a copywriter charging per word, you should be charging per second." Call-outs work because they let the viewer self-identify in under a second, and self-identified viewers watch longer. They die the moment you broaden the audience: "If you want to grow on social media" addresses no one. Tighten the call-out until it cuts someone out.

The contrarian claim. "90% of TikTok hook advice is for people who already know how to write hooks." The contrarian claim is the most overused pattern in 2025, which means it's also the easiest to get wrong. The rule: your contrarian claim has to be specific enough that half the audience disagrees. "TikTok is oversaturated" is a claim nobody fights. "TikTok hooks are easier to write than email subject lines, and here's why" is one a quarter of the audience will scroll past in fury — and that scroll-past is the point.

One pattern I avoid: the rhetorical question. "Want to know the secret to going viral?" worked in 2019. By 2026 it reads as a 404 page. If you find yourself starting a hook with "Do you want to know…," rewrite it as a statement.

The AI workflow that actually ships

This is the part most "AI hook generator" articles get wrong. They show you a prompt that returns ten generic hooks, and call that the workflow. That's not a workflow — that's a content spinner.

The workflow I run has four steps, and only one of them involves the AI drafting.

Step 1: Feed Claude my last 10 videos with their retention curves. Before I ask it to write anything, I paste in the transcripts of my ten most recent posts along with the second-3-second drop-off rate from TikTok analytics. I ask: "Based on the transcripts, what linguistic features appear in the videos where under 40% of viewers dropped off in the first 3 seconds, and what features appear in the ones where over 70% dropped off?" This step takes 5 minutes and gives me the model my style, not the average TikTok creator's style.

Step 2: Have it draft 20 hooks for the new script. Here's the prompt I actually use, with a placeholder:

Here are 5 transcript excerpts from my own videos that retained well in the first 3 seconds: [paste]. I'm about to film a TikTok on [topic], with this working title: [title]. Draft 20 opening lines for the first 3 seconds, in the voice and style of those excerpts. For each line, label which of the 5 patterns from my earlier analysis it uses. Vary the patterns — don't give me 20 versions of the same one.

I almost always get 20 usable lines back. I never publish any of them.

Step 3: I personally rank them, blind, and kill 16. I paste the 20 into a Google Doc, scramble the order, and rank them 1–20 without knowing which pattern each one uses. The pattern is irrelevant to my ranking — I just go with gut. Then I cut everything below #4. Four is the right number; it gives me variety on set without spreading my filming energy thin.

Step 4: Film all four, learn from the one that wins. This is the step the AI workflow articles always skip. AB testing isn't optional — it's the only way to know which hook style is working for your audience right now. Once a month I go back through the four I've shipped, look at the 3-second retention, and update Step 1's transcript pile.

The AI's job in this workflow is to be a fast, stylistically aware brainstormer. My job is to be the editor. If you skip the editing step, you ship a generic AI video. If you skip the filming step, you don't actually know which hooks work.

A real before-and-after

A bad hook, the kind I would have written in 2021:

"Hey guys! In today's video, I'm going to share 3 tips for better marketing with AI. Let's get into it!"

Four problems in two sentences: it opens with a generic greeting, the body is a list, the language is corporate ("share 3 tips"), and there's no reason to watch the rest — you've already told me what's in the video.

The rewrite, given the same script about AI marketing tips:

"I spent $4,200 on AI tools last year. Two of them changed my workflow. The other nine are basically lamps."

That hook does all three jobs. The number is specific. The genre is set in one phrase (this is a personal take, not a tutorial). The loop — which two, and why — is open by the end of the first sentence, and the joke at the end is just enough texture to make me trust the speaker.

The body of the video is identical between the two versions. The hook is doing all the work.

The kill list

Things I cut from AI-drafted hooks before I'll film them. None of these are controversial; they're the easy giveaways that a line was written by a model that has read too many "viral post" blog posts.

• "Hey guys," "what's up," "in today's video," "let's get into it," "without further ado." All of them are zero-information openers.
• Adverbs: really, actually, literally, just. The model is addicted to them. So are most creators, which is why they're a red flag.
• "I'm going to show you" / "let me show you." The viewer is on a phone, not in a classroom.
• Any hook that takes more than 6 words before the main claim.
• Em-dashes mid-sentence. I have a private theory that em-dashes in short-form video read as LinkedIn. The data doesn't support it; the gut does.
• A question that takes a full sentence to set up. "Have you ever wondered why some videos go viral and others don't?" has used 12 words to say nothing.

If a hook survives the kill list, I read it out loud. If it sounds like a person talking to another person, it ships. If it sounds like a content marketer reading a brief, it doesn't.

The frame that makes the work mechanical

The hook is the part of the video the algorithm decided to test, not the part the viewer chose to watch. Once you frame it that way, the work stops being creative and starts being procedural: write 20, kill 16, film 4, learn from the one that wins. The first 3 seconds isn't where creativity lives. It's where taste lives — and taste is a thing you build by shipping 200 mediocre hooks and remembering which 20 felt right.

The algorithm isn't a mystery. The hook isn't a vibe. It's a job. Do the job.