The biggest failure mode of AI-written short-video scripts isn’t bad writing. It’s that they read fine on the page and die when spoken aloud. AI drafts in essay pace: full sentences, polite transitions, zero camera awareness. And in 2026 the cost of a slow open is brutal. TikTok’s Q2 2026 algorithm weights the 3-second hold above almost everything else, and videos that hold 60-65%+ of viewers past second three pull roughly 4-7x more impressions than videos that leak viewers immediately (per OpusClip and Sprout Social retention studies). This workflow gives you a 6-part prompt brief, a 5-hook rule, and the read-aloud trim that turns AI prose into something you can actually film.
TL;DR
Brief any frontier model (ChatGPT, Claude, or Gemini) with six specifics, force 5 hook variants and reject the first one, generate a structured 30-90s script, read it aloud and cut ~20%, then add inline direction notes the AI never writes. The whole pipeline drops from ~60 minutes on script one to 15-20 minutes once you save templates.
Platform length and pacing cheat sheet (June 2026)
Match the script length to where it ships. Maximum length and best-performing length are not the same number, and the gap matters because every platform’s algorithm rewards completion rate.
| Platform | Hard max (2026) | Best-performing length | Pacing |
|---|---|---|---|
| TikTok | 10 min in-app, 60 min uploaded | 15-30s | Fast cuts, audio-led hook |
| Instagram Reels | up to 3 min in-app (longer uploads exist) | 15-45s; viral loop zone 7-15s | Mid-pace, strong loop |
| YouTube Shorts | 3 min (since Oct 2024) | 20-45s | Mid-pace, end-to-end watch |
| LinkedIn video | ~10 min | 30-60s | Slower, caption-readable |
YouTube extended Shorts to 3 minutes in October 2024, and Instagram now records up to 3 minutes in-app, but a 20-second Short at 90% completion still buries a 3-minute one at 15%. Note on music: any Short over 1 minute with an active Content ID claim is blocked globally, so keep licensed tracks under 60-90 seconds.
Pick a model for the job
You do not need a paid plan to start. All three frontier assistants have free tiers that draft a usable script, but they have different strengths for this task.
| Model (June 2026) | Free tier | Best for short-form | Notes |
|---|---|---|---|
| ChatGPT (GPT-5.5) | Yes (tight limits, US Free has ads) | Punchy 3-second hooks, tight rewrites | Plus $20/mo lifts limits; picker has Instant / Thinking / Pro |
| Claude (Sonnet 4.6) | Yes (limited) | Natural spoken voice, tone matching, fewer robotic transitions | Pro $20/mo ($17 annual) |
| Gemini 3.1 Pro | In Google AI Pro $19.99/mo | Live web facts in the hook, longer narrative arcs | Renamed from “Gemini Advanced” early 2026; 1M-token context |
A practical combo: draft hooks in ChatGPT (GPT-5.5), then have Claude (Sonnet 4.6) rewrite the chosen script for natural delivery. If you are still deciding between the three, see our ChatGPT vs Claude vs Gemini comparison.
Before you start
- Pull 3-5 reference videos in the style you want and paste the URLs or transcripts into the session. “Match this energy” with examples beats any amount of prose description.
- Decide the platform first. Reels and Shorts sit at 15-45s with mid pace; TikTok rewards 15-30s with faster cuts. Hook conventions differ too.
- Name your goal: entertain, educate, or sell. Every model defaults to “educate” and produces a dry voice unless told otherwise.
- Have the CTA ready before scripting. “Follow for part 2” / “Comment the keyword” / “Link in bio”. The script should end on the CTA, not invent it at the last second.
Step by step
- Brief with 6 components. Niche / target viewer, platform and length (Reels 30s / TikTok 20s / Shorts 30s), goal (entertain / educate / sell), the one-sentence promise the video delivers, your voice in one adjective, and the CTA. Skip any of these and the model guesses badly.
- Generate 5 hook variants. Never use the first hook the model writes; it is almost always the safest and weakest. Specify patterns that actually score above 70% retention in 2026: Identity Call (“If you sell on TikTok Shop, this is for you”), Contrarian Strike (“Stop posting 60-second videos”), Open Loop (“This $14 tool replaced my $40 one”), Confession (“I lost $4,000 before I learned this”), and a specific-number hook.
- Pick 1 hook and generate the full script in the structure: hook (3s) → middle (45-60s of value or story) → payoff (10s of insight or resolution) → CTA (5s).
- Read it aloud and time it. AI scripts run roughly 20% long. Cut by reading and trimming any word that doesn’t change the meaning when removed.
- Add direction notes inline:
cut on word X,B-roll: phone hand-holding,text overlay on word Y,zoom-in on word Z,music drops here. Layered hooks (visual + audio + on-screen text) lift 3-second holds by up to 3x versus a single-element open, so plan all three layers in the first two seconds. The model never adds these. You must. - Build a shot list from the direction notes. This is what you look at while filming; the script is for the audio track.
- Film, edit, post. Save the winning hook and script as a template. The next script in this style takes 15 minutes instead of 60.
A worked example: a 25-second TikTok
Brief: niche = indie SaaS founders, platform = TikTok 25s, goal = educate-to-sell, promise = “cut your onboarding emails in half”, voice = blunt, CTA = “Comment ‘ONBOARD’ for the template.”
Hook (Contrarian Strike, 3s): “Stop writing five onboarding emails. You need two.” rapid zoom-in on word 'two' text overlay: 2 > 5
Middle (16s): the two emails that matter, why the other three get ignored, one concrete open-rate number you actually measured. B-roll: inbox scroll cut on each email number
Payoff (3s): “The whole sequence took me one afternoon to rewrite.” talking head, slow zoom
CTA (3s): “Comment ‘ONBOARD’ and I’ll send the template.” text overlay: ONBOARD ⬇️
Notice the script is ~70 spoken words. That is the right density for 25 seconds. AI will hand you 95+ words for the same slot; the read-aloud trim is where you get back to filmable.
Quality check
- Does the hook land in 3 seconds? Read only the first sentence aloud. If it’s setup rather than payoff, the hook is too slow, and the algorithm punishes a slow open harder than ever in 2026.
- Does the script speak well? Sentences over 12 words die in delivery. Use contractions everywhere. Strip “however” and “additionally”.
- Does each line earn its place? Cut any line that can be removed without losing meaning.
- Are direction notes in the script? Otherwise you improvise on set and your editor improvises in post. Both cost quality.
- Does the CTA fit the platform? “Subscribe” on TikTok converts worse than “follow for part 2” or “comment the keyword”.
How to reuse this workflow
- Save the 6-component brief per content pillar. “Educational mini-explainer”, “behind-the-scenes story”, and “before-after demo” each get their own brief.
- Build a hook library by pattern. After ~20 scripts you’ll know which of the four high-retention patterns your audience responds to and can skip the losers.
- Save winning scripts as templates with one variable (topic / number / claim) to swap. A weekly batch of 5 scripts becomes ~30 minutes.
- Re-check platform norms every quarter. TikTok shifted its weighting toward 3-second retention in Q2 2026, and length conventions keep moving.
Common mistakes
- No hook variants. The first hook the model writes is almost never the best. Always generate 5.
- Scripts that read but don’t speak. Every model defaults to essay pace. Read aloud and cut.
- No direction notes, so your editor (or future you) forgets the intended timing of cuts, B-roll, and text.
- Generic CTAs like “smash that like button”. Pick a CTA that fits the platform.
- Ignoring 3-second retention data. It is your hook’s report card; use it to retire weak patterns.
- One length for every platform. A 30s Reel and a 15s TikTok need fundamentally different pacing.
FAQ
- What’s the right script length per platform in 2026?: Best-performing zones are TikTok 15-30s, Reels 15-45s (viral loop zone 7-15s), Shorts 20-45s, and LinkedIn 30-60s. Hard maximums are far higher (Shorts and in-app Reels go to 3 minutes), but completion rate falls off fast past ~45 seconds.
- Which AI is best for short-form scripts?: ChatGPT (GPT-5.5) writes the punchiest 3-second hooks; Claude (Sonnet 4.6) produces the most natural spoken voice; Gemini 3.1 Pro pulls live facts into the hook. Free tiers of all three are enough to start.
- How do I make AI sound like me?: Feed it 3-5 of your best past scripts as reference and say “write in this voice.” Without examples, every model defaults to a generic influencer voice.
- Should I use AI voiceover?: For utility lines, sure. For the main delivery your real voice still outperforms; audiences clock an AI voice in under two seconds.
- How long until the workflow is fast?: First script: about 60 minutes. After 10 scripts with templates: 15-20 minutes each.