TL;DR
About half of viewers drop off in the first 3 seconds of a Reel, so the opening line is the single highest-leverage thing you can rewrite. Hooks are a numbers game: generate 20 candidates with AI, post one variant at a time, and keep the 2-3 that push your 3-second hold rate above 60%. Use Claude Sonnet 4.6 for natural, spoken-sounding hooks and GPT-5.5 when you want punchier marketing angles. The prompt below is the working part of this page — copy it, fill the five bracketed slots, and run it.
Why the first 1.5 seconds decide everything
Instagram ranks Reels mostly on watch time. In January 2025, Adam Mosseri confirmed the three signals that matter most on the Reels surface: watch time, likes per reach, and sends per reach (a DM share is weighted roughly 3-5x heavier than a like for reaching new accounts). All three collapse if people never get past the hook.
The data is brutal. Up to 50% of viewers scroll away in the first three seconds, and Reels with a 3-second hold rate above 60% out-reach those below 40% by 5-10x, per creator-analytics aggregates as of June 2026. The caption overlay doesn’t even render until ~1.5 seconds in, so your spoken opening has to do the work alone. Most creators write one hook, post it, then wonder why one Reel does ten times the views of another. The honest answer is almost always the hook.
That is exactly the kind of problem AI is good at: producing 20 candidates fast, so you can spend your own energy editing the three that survive instead of staring at a blank line.
The numbers worth optimizing for
| Metric | What it measures | Healthy target (June 2026) |
|---|---|---|
| 3-second hold rate | % who watch past 3s | Above 60% |
| Average watch time | Total seconds viewed | The single top Reels ranking signal |
| Sends per reach | DM shares ÷ accounts reached | Most heavily weighted for new-audience reach |
| Reel length | Length of the clip | 7-30s converts best for most niches |
You cannot edit “sends per reach” directly. You edit the hook, the hook lifts hold rate and watch time, and Instagram responds by widening distribution. The hook is the lever; everything downstream is the result.
One nuance for 2026: Instagram is watching for AI slop
In a year-end memo on December 31, 2025, Mosseri said Instagram would prioritize “raw, real human content” over obviously AI-generated material through 2026. That does not mean you cannot use AI to brainstorm. It means the output must not sound generated. Treat the model’s 20 lines as raw clay. The final hook should be edited by you, in your voice, anchored to one specific image or claim that only you would write. AI hooks that ship unedited drift toward the safe middle and sound like every other Reel, because they were trained on every other Reel.
What to feed the model
Garbage in, generic out. Give it five things:
- The exact topic of the Reel (specific beats broad: “pricing your first freelance logo” not “freelancing tips”)
- The audience — who they are, what they already know, what they fear
- The fear or desire driving them to watch
- Your tone — calm, hyper, contrarian, warm
- 1-2 of your own past hooks that performed, as a voice reference
Copy-ready prompt
Paste this into Claude (Sonnet 4.6) or ChatGPT (GPT-5.5). Replace each bracketed slot with your own text. Do not leave the brackets in.
Generate 20 Instagram Reel hooks.
Topic: [topic]
Audience: [audience] # e.g. "freelance designers, 2-5 years in"
Audience fear or desire: [pain or desire]
My voice (sample): [paste 1-2 of my past hooks]
Hook rules:
- Each hook is one sentence, max 12 words.
- Each must be speakable inside the first 1.5 seconds of the video.
- No "Hey guys" / "In this video" / "Welcome back" openers.
- Vary the type. Use exactly 4 of each:
1. Pattern interrupt ("Stop scrolling. Your X is wrong.")
2. Direct question ("Why does your X keep doing Y?")
3. Contrarian claim ("Everyone says X. They are wrong.")
4. List tease ("3 things I wish I knew before X.")
5. Story open ("A client told me X yesterday. Here is what I said.")
Output a numbered list grouped by type. No preamble.A flat list grouped by type makes A/B testing trivial: pull one hook from each of the five categories for your first week of posts, and let hold rate tell you which structure lands for your account.
Which model to use
Both work; they fail differently.
- Claude Sonnet 4.6 tends to write hooks that sound like a person mid-thought — incomplete sentences, natural verbal tics. That reads better for Reels, where authenticity drives completion. It also follows multi-rule prompts more reliably (roughly 14 of 15 constraints held in head-to-head tests vs. 11-12 for ChatGPT), so the “max 12 words / 4 per type” rules stick.
- GPT-5.5 writes tighter, more traditionally “marketing” hooks. Great when you want a punchy pattern interrupt, occasionally too clever to sound human.
A practical move: run the same prompt through both and merge the best of each list. As of June 2026, Claude Pro is $20/mo and ChatGPT Plus is $20/mo; both free tiers can run this prompt, just with tighter daily limits.
How to vet the 20 lines
- Read each one out loud at normal speaking pace. If it doesn’t fit in ~1.5 seconds, cut it.
- Delete anything using jargon your audience would not say aloud.
- Star the 3 that genuinely surprised you. Those are usually the winners.
- Kill any hook that promises something the rest of the Reel doesn’t deliver. A broken promise tanks watch time and trains the algorithm to bury you.
Common mistakes
- Testing one hook style (usually questions). Variety is the entire point of generating 20.
- Declaring a winner after two posts. You need at least 5 posts per hook style before a pattern is real; plan 8-12 posts across 2-4 weeks.
- Posting the raw AI line. Edit it into your voice or it sounds like slop, the exact thing the 2026 algorithm is told to demote.
- Ignoring what’s trending this week. AI doesn’t know what hit Instagram today. Add 2-3 hooks you scrolled past this morning to your test batch.
Keep a swipe file
Save every hook from your top 10 best-performing Reels in one note. Paste those back into the prompt’s voice-sample slot to anchor future generations in what already works on your account, not generic Reel-speak. Over a few months this becomes the most valuable input you give the model.
FAQ
- How many hooks should I actually test? Plan for 8-12 posts across 2-4 weeks, at least 5 per hook style, before drawing conclusions. Fewer than that is noise.
- What hold rate should I be aiming for? Above 60% at the 3-second mark. Below 40% and the Reel is effectively dead on arrival. Check it in Instagram Insights under each Reel’s retention graph.
- Should the on-screen text and spoken hook match? Usually yes; reinforcement helps retention. Split them only when the spoken line would look awkward as caption text.
- Will Instagram penalize me for using AI to write hooks? No. The 2026 “raw human content” push targets obviously AI-generated video and voices, not your brainstorming process. A human-edited hook in your own voice is fine.
- Claude or ChatGPT for this? Claude Sonnet 4.6 for natural, spoken-sounding hooks; GPT-5.5 for punchier marketing angles. Run both and merge if you can.
Related
- Prompt library: reel hook prompts
- Prompt library: reel caption prompts
- Short-form video hook with AI
External: Adam Mosseri on how Reels are ranked and Instagram’s in-app Insights retention graph are the two sources worth checking before every test cycle.