TL;DR
Your first 3 seconds decide whether 60-80% of viewers stay or scroll. Feed an AI the video’s concrete payoff plus your audience and platform, then make it write 8 hooks from 8 different pattern families (one each). Pick 2 from different families, film both, and A/B them. The number you’re chasing: keep three-second retention above ~65% — TikTok for Business reports 63% of the highest click-through videos hook inside the first three seconds. AI is reliable at applying hook patterns; it cannot tell you which pattern wins in your niche. That part is empirical.
The task
You have a 38-second TikTok edited and ready, but the first 3 seconds — the bit that decides whether most viewers stay or scroll — is a placeholder you wrote tired. Your retention curve from the last 5 videos is a cliff: 100% at second 0, 32% at second 3, 20% by second 6. The content is fine; the hook is leaking 68 points of audience before anyone sees the payoff. You want 8 hook variants from different pattern families so you can pick two to actually film, test against each other, and learn what works for your specific niche over the next two weeks.
The stakes are concrete. As of June 2026, the completion-rate bar for viral distribution has risen to roughly 70% (it was around 50% in 2024). A weak hook doesn’t just cost the current video — it drags down completion rate, which is the single strongest ranking signal TikTok, Reels, and Shorts feed into distribution.
What “good” looks like (June 2026 benchmarks)
These are the numbers to test your hook against, not vanity view counts.
| Metric | Target / benchmark | Source context |
|---|---|---|
| 3-second retention | Above ~65% (top videos hit 70-85%) | TikTok for Business hook research |
| Completion rate for viral push | ~70%+ | Up from ~50% in 2024 |
| Avg watch time, video under 30s | Above ~50% is solid | 2026 retention benchmarks |
| Avg watch time, 30-60s video | Above ~40% | 2026 retention benchmarks |
| Completion by length | ≤10s ≈ 81%, 31-60s ≈ 42%, 60s+ ≈ 29% | Length drives the ceiling |
| Sound-off viewers | ~80% across platforms | Caption the hook, every time |
Two takeaways the table forces. First, if you can land the same payoff in 21-34 seconds you start from a much higher completion ceiling, so shorter edits forgive a weaker hook. Second, because roughly 80% of viewers watch on mute, an audio-only hook reaches one in five people in the first 3 seconds — on-screen text is not optional.
Where AI helps — and where it does not
AI is good at applying hook patterns reliably: counter-intuitive claim, stake/consequence, concrete number, personal admission, open question, “if you” filter, before/after tease, mid-action time-stamp. It can also help you avoid patterns that went invisible by 2026 (“have you ever,” “today I’m going to show you,” “you won’t believe”).
What AI cannot do: tell you which pattern works in your niche. That is purely empirical and varies by audience — beauty creators often see counter-intuitive hooks underperform; finance creators see them overperform. Test 2 hooks from different pattern families against each other for at least 2 weeks before settling on a default. AI also cannot judge whether your hook overpromises against your actual payoff. If the hook claims something the video doesn’t deliver, watch-time tanks and the algorithm penalizes the next post too.
A specific failure mode: AI gravitates to grandiose opening words (“incredible,” “shocking,” “you won’t believe”). Tell it plainly: the first 3 syllables must be unexpected but plain — no superlatives, no clickbait register. GPT-5.5 and Claude Sonnet 4.6 both default to this hype register unless you forbid it in the prompt; the rules block below does that.
What to feed the AI
- The actual payoff of the video — what the viewer walks away with (not the topic, the concrete takeaway)
- Audience level — tourist (sees you for the first time), beginner (follows the niche), regular (knows your style)
- The platform — TikTok, Reels, Shorts each have slightly different rhythm and tolerance for jump-cuts
- One specific word or phrase to avoid (one that’s overused in your niche right now)
- Your retention curve from recent posts — where do people drop? If they drop at second 5, the hook needs an open loop that pays off after second 5
- The hook formula your top 3 best-performing videos used (the model will pattern-match it)
- Whether the hook will be on-screen text, voice-over, or both
- A note on tone — dry / earnest / playful / sarcastic — and one creator whose hook voice yours most resembles
Copy-ready prompt
Paste this into ChatGPT (GPT-5.5), Claude (Sonnet 4.6), or Gemini 3.1 Pro. Replace each [bracketed] placeholder.
Write 8 short-form video hook candidates. Each hook 1-2 sentences, total under 4 seconds spoken.
Video payoff (concrete takeaway): [paste]
Audience: [tourist / beginner / regular]
Platform: [TikTok / Reels / Shorts]
Tone: [dry / earnest / playful / sarcastic] — like [creator name]
Avoid this word / phrase: [paste]
Hook formula my last hit used: [paste]
Hook delivery: [on-screen text / VO / both]
Use these 8 pattern families, one each:
1) Counter-intuitive claim — a sentence that contradicts the niche's default belief
2) Stake / consequence — what the viewer loses by not watching, in concrete terms
3) Concrete number — a specific, slightly odd number ("3.5 hours," "47%," "11 out of 30")
4) Personal admission — something you got wrong, lost, or were embarrassed by
5) Open question — a question the viewer cannot answer themselves; payoff must land in the video
6) "If you ___, this is for you." — a targeted filter the right viewer recognizes
7) Before/after tease — show the after-state in 3 seconds, promise the path
8) Mid-action timestamp — "Day 3 of testing every AI ___" — implies a series and creates a loop
Rules:
- First 3 syllables must be unexpected but plain. No superlatives ("incredible," "amazing," "shocking"). No clickbait register.
- Each hook must be deliverable in under 4 seconds spoken.
- Do not promise a payoff the video doesn't deliver.
- If using on-screen text + VO, give me the 5-word on-screen line AND the 8-word VO that adds (not repeats).Shorter variant — rapid iteration on a specific hook
Below is a hook that under-performed (retention dropped from 100% to 28% by second 3). Rewrite into 5 variants that fix specifically what makes it weak — be explicit about what you changed and why. Use no superlatives.
Hook: [paste]
Video payoff: [paste]Sample output
A useful counter-intuitive hook: “I stopped writing prompts longer than 100 words — my output got measurably better.” (On-screen text: “Short prompts beat long ones.” VO adds the “measurably” detail.)
A useful stake hook: “If your AI keeps making up citations, you have one minute to check this — the fix is in your settings, not the prompt.”
A useful personal-admission hook: “I spent 8 months building the wrong AI feature. Here’s the question I should have asked first.”
A useful mid-action hook: “Day 3 of testing every AI photo editor — and the one that won is the one no creator I know talks about.”
A useful “if you” filter: “If you’re using ChatGPT for research and not citing sources, this 40-second clip will save you a Reddit pile-on.”
Take the hook straight into your edit
The hook text the AI gives you has to become animated, synced captions, or sound-off viewers see nothing. As of June 2026 the practical chain is:
- CapCut — free, the most common option; auto-captions sync to speech and let you punch up the emphasis word during the hook. Best default if you edit on phone or desktop.
- OpusClip — repurposes a long video into shorts and writes the captions for you (free plan ~60 credits/month, Starter $15/mo, Pro $29/mo as of June 2026). Useful when the hook is being cut out of an existing recording.
- Captions.ai — still strong for mobile-first caption styling.
Whatever you use, keep the hook line in the top third of the frame: the bottom is hidden by the comment overlay and the username strip, and display 2-3 words at a time so a muted viewer can read it at a glance. For the surrounding script, see Write a Short Video Script With AI; for finding the angle in the first place, AI Short Video Ideation.
How to refine
- Strip the superlatives: “Remove every ‘incredible,’ ‘amazing,’ ‘shocking,’ ‘you won’t believe,’ or ‘mind-blowing.’ Rewrite to land the same surprise without superlatives. Plain-but-unexpected is the goal.”
- First 3 syllables must be unexpected: “Re-read each hook. If the first 3 syllables are ‘today I’m,’ ‘have you ever,’ or ‘let me show you,’ rewrite. The first word should be a noun or a verb the viewer did not predict.”
- Open a loop the video closes: “Each hook must implicitly promise something specific the video delivers — and the video must deliver it within the first 30 seconds. If the hook would still work with a different video, it is too vague.”
- Test across pattern families, not within: “Pick 2 hooks from 2 different families (counter-intuitive + stake). Two counter-intuitive variants against each other tell you which sentence is sharper; testing across families tells you what your audience responds to.”
- On-screen text and VO add, not repeat: “If hook is on-screen text + VO, give me 5-word on-screen and 8-word VO. The VO must add a detail not in the text; if it repeats, cut the VO and let the visuals carry the next 2 seconds.”
Common mistakes
- “Have you ever” or “today I’m going to show you” — universally weak in 2026; viewers recognize the pattern and pre-scroll
- Hooking with content the video doesn’t deliver — burns audience trust, and the algorithm penalizes the next post too
- Using the same hook style on every video — the algorithm reads it as signal staleness and underdistributes; vary across pattern families
- Reading the hook into camera with no on-screen text — about 80% of viewers watch on mute, so they get nothing in the first 3 seconds
- Hooks longer than 4 seconds spoken — by second 4 the algorithm has already measured “stayed past 3”; you are spending engagement budget on hook, not payoff
- Superlatives in the first sentence (“This is the most insane AI tool”) — clickbait register gets filtered by both algorithm and viewers
- Burying the on-screen hook below the action — the bottom of the frame is hidden by the comment overlay; keep hook text in the top third
- Posting one hook and concluding — always batch-record 2 hooks for the same edit; A/B reveals signal you cannot guess
FAQ
- Should the hook be on-screen text or VO?: Both, when possible. With roughly 80% of viewers on mute, on-screen text is the floor, not the bonus. Strongest combo: a 5-word on-screen line plus an 8-word VO that adds context rather than repeats it. Keep the text in the top third of the frame.
- What three-second retention should I aim for?: Above ~65% as of June 2026; the videos that travel sit at 70-85% at second 3. If you’re under 50% at second 3, the hook — not the content — is the problem to fix first.
- Do these patterns work for LinkedIn video and YouTube Shorts too?: Yes, with adjustments. LinkedIn: cut the “personal admission” pattern — it skews oversharing on that platform. Shorts: lean harder into the “if you” filter, because Shorts viewers are more search-led than feed-led.
- How many hooks should I test before settling on a default?: At least 2 per post for 2 weeks (roughly 8-14 hooks across families). Compare retention at second 3 and second 8 across the cohort, not absolute view counts. Hooks that pass second 3 but fail second 8 are clickbait.
- Can I reuse a hook that worked once?: Once or twice, on different content. By the third time, viewers recognize it and the algorithm reads repetition and underdistributes. Refresh the language, keep the pattern family.
- The model keeps writing clickbait — what changes?: Add: “No superlatives. The hook must be plain but unexpected. If a hook needs a word like ‘incredible’ to work, the underlying claim is the wrong claim and the hook should be rewritten around a different angle.” GPT-5.5 and Sonnet 4.6 both respond to this constraint reliably.