The task
You have a 45-second TikTok edited and ready, but the first 3 seconds — the bit that decides whether 80% of viewers stay or scroll — is a placeholder you wrote tired. Your retention curve from the last 5 videos is hockey-stick: 100% at second 0, 32% at second 3, 20% by second 6. The video content is fine; the hook is leaking 70 points of audience before they ever see 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.
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 the patterns that have become invisible in 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 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 punishes the next post too.
A specific failure mode: AI gravitates to grandiose opening words (“incredible,” “shocking,” “you won’t believe”). Tell it: “the first 3 syllables must be unexpected but plain — no superlatives, no clickbait register.”
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 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/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, VO, or both
- A note on tone — dry / earnest / playful / sarcastic — and one creator whose hook voice yours most resembles
Copy-ready prompt
Write 8 short-form video hook candidates. Each hook 1-2 sentences, total ≤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 loop
Rules:
- First 3 syllables must be unexpected but plain. No superlatives ("incredible," "amazing," "shocking"). No clickbait register.
- Each hook must be deliverable in ≤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.”
How to refine
- Strip the superlatives: “Remove every ‘incredible,’ ‘amazing,’ ‘shocking,’ ‘you won’t believe,’ or ‘mind-blowing.’ Rewrite the hook 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 actually 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 pattern families (e.g., counter-intuitive + stake) for the next post. Testing 2 counter-intuitive variants against each other tells 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 delivered as 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 does not actually 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 the camera with no on-screen text — captions-off viewers (about half on TikTok and Reels) get nothing in the first 3 seconds
- Hooks longer than 4 seconds spoken — by second 4 the algorithm has already measured “stayed past 3” and you are spending engagement budget on hook, not payoff
- Superlatives in the first sentence (“This is the most insane AI tool”) — clickbait register is filtered by both the 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 video edit; A/B reveals signal you cannot guess
FAQ
- Should the hook be on-screen text or VO?: Both, when possible. Strongest combo: 5-word on-screen text + 8-word VO that adds context, not repeats. Captions-off viewers get the text; sound-on viewers get the layered detail. Top-of-frame for text.
- 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 pattern, 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 (so roughly 8-14 hooks across pattern families). Look at retention at second 3 and second 8 across the cohort, not view counts in absolute terms. Patterns that pass second 3 but fail second 8 are clickbait hooks.
- Can I reuse a hook that worked once?: Once or twice, on different content. Three times and viewers recognize it; algorithm reads it as repetition and underdistributes. Refresh the language while keeping 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.”