TL;DR
The first line decides whether anyone watches or scrolls past. Feed your AI one topic plus four inputs (niche, audience fear/desire, platform, and any real number you own), and get back 20 opening lines tagged by hook pattern and ranked by stopping power. Then test three with different patterns — not one. The prompt is below, with the platform length limits and the retention bar your hook actually has to clear.
Why use AI for this, and where it breaks
A human writer usually defaults to one or two hook patterns they personally like. AI’s edge is coverage: ask once and it lays curiosity-gap, contrarian, vivid-scene, numeric, social-proof, and pattern-interrupt variations side by side on the same topic. That horizontal spread is exactly how you find a winner you would never have written yourself.
Where it breaks: AI will invent specifics if you let it. If a hook says “I tripled revenue in 30 days,” that number has to come from you. Fabricated claims draw comment-section corrections and depress watch time, which on AI-curated feeds is the metric that decides distribution. Treat the model as a phrasing engine, never a fact source — the prompt below hard-blocks any number you did not supply.
The four inputs that decide quality
Thin inputs produce generic hooks. Give the model these every time:
- Niche — be narrow: “personal finance for new grads,” not “money.”
- Topic — what this one post is actually about.
- Audience fear or desire — one specific thing they secretly want or dread.
- Platform — line length and rhythm differ by feed (see the table below).
Copy-ready prompt
You are a short-form copywriter who specializes in opening lines that
hold viewers past the 3-second mark.
Generate 20 opening lines for a [platform] post about [topic] in the
niche [niche]. Audience pain or desire: [pain_or_desire].
Rules:
- Each line under 12 words, conversational, easy to say out loud.
- Label each line with ONE pattern: curiosity-gap, contrarian,
vivid-scene, numeric, social-proof, pattern-interrupt, listicle,
callout, or question.
- Do not include any specific number, stat, or claim I did not supply.
- No "Hey guys," no "Check this out," no logo/intro framing.
Rank all 20 strongest to weakest by stop-the-scroll power and return a
numbered list. Below the list, name the 3 lines (with different patterns)
you would A/B test first and say why in one sentence each.Swap the bracketed [placeholders] for your real inputs. The “different patterns” instruction matters: you want your three tests to disagree on approach, so the winner tells you something.
Platform length limits (as of June 2026)
Where your hook gets cut off is non-negotiable, and it is not the same everywhere. Design for the fold:
| Platform | Hook budget | What cuts it off |
|---|---|---|
| TikTok / Reels / Shorts | ~7 words of on-screen text; spoken hook in first 2-3 sec | The 3-second retention window, not a character count |
| X (free) | 280 characters per post | Hard post limit; Premium long posts still preview at ~280 in-feed |
| ~140 characters on mobile, ~210 on desktop | The “…see more” fold; most traffic is mobile, so write for 140 |
For video, the constraint is time, not characters: the spoken or on-screen hook must land in the first 2 to 3 seconds, because that window drives the bulk of completion-rate variance on AI-ranked feeds.
The retention bar your hook has to clear
Hooks are not judged on vibes — short-form feeds measure them. Use these as your pass/fail line when you check published results:
| 3-second retention | Read it as |
|---|---|
| Above ~65% | Strong hook; videos here reportedly earn several times more impressions |
| 50-65% | Workable, keep iterating the first line |
| Below ~50% | Weak hook — rewrite the opening line, not the video |
Losing more than about a third of viewers in the first three seconds is the signal a hook is failing. That is why testing three opening lines on the same clip is the fastest fix: you isolate the one variable that moves the number.
How to vet the output before you post
- Read the top five aloud. If you have to breathe before the period, it is too long for short-form — cut it.
- Delete any line that needs niche jargon to land; people without context bounce.
- Check every number against your inputs. If you did not supply it, delete the line even if the prompt was supposed to block it.
- Lead with conflict or payoff, never context. An early logo card or “Hey guys” reads as low-engagement to the feed and suppresses reach.
Common mistakes
- Click-bait with no payoff — the hook promises what the body cannot deliver, and the drop-off shows up in your retention graph.
- Keeping every variant — pick three with different patterns, kill the rest, test.
- One hook across every platform — a LinkedIn line at 140 characters is a different object than a 2-second spoken Reel hook. Retune length and rhythm per feed.
- Stacking only one trigger — the strongest hooks combine a pattern interrupt, a curiosity gap, and a payoff signal in one breath.
Find your hit pattern
Track which pattern wins on each platform for one month. Most creators discover a personal “hit pattern” — curiosity-gap for some, contrarian for others. Once you know yours, bias future prompts toward it: ask the model for fresh variations inside your winning pattern instead of another wide horizontal sweep.
FAQ
- How many lines should I actually test? Three, each a different pattern, on the same topic and ideally the same clip, within one week. One pattern wins and becomes your default.
- What retention number means my hook worked? On short-form video, aim for above ~65% retention at the 3-second mark; below ~50% means rewrite the opening line. (Benchmarks vary by niche and platform.)
- Does this work for long-form articles? Yes — same prompt, but ask for sub-heads instead of opening lines and allow up to ~15 words each.
- Which model should I run this on? Any current frontier model handles it well — GPT-5.5, Claude Sonnet 4.6, or Gemini 3.1 Pro. The output quality tracks your inputs far more than the model choice.
- Every variant still feels generic — why? Your input is too thin. Add a one-line true story or a specific number you actually own, then re-run.