Good AI prompts for TikTok and Instagram Reels scripts give the model your audience, the single point of the video and the length, then ask for several hooks, a shot-by-shot script with on-screen text, and a spoken version timed to the video length. The opening seconds deserve the most attention: TikTok’s own recommendation page says watch time is generally weighted more heavily than other signals, and its ad guidance advises hooking viewers early. AI is good at producing hook variations and tightening a script; it is bad at knowing what your audience finds surprising, so you supply that part.
This guide covers what the platforms actually say about short videos, a four-prompt workflow you can copy, a worked example from idea to shot list, and the mistakes that make AI-written scripts sound like AI.
What the platforms say about the first seconds
The case for obsessing over the opening comes from the platforms themselves, not from folklore about a magic three-second rule. TikTok’s help page on how For You recommendations work lists whether people watch a video in full or skip it among the signals it uses, and says time spent watching is generally weighted more heavily than other factors for many users. If viewers leave in the first moments, the video has little chance to be shown more widely.
TikTok’s creative best practices for performance ads are written for advertisers, but they are the most concrete official guidance available and translate well to organic content. Paraphrased, they recommend:
- putting the hook in the first six seconds, using suspense, surprise or another emotional pull;
- introducing what the video is about within the first three seconds;
- shooting vertically at 9:16 with sound;
- adding captions or text overlays for context, at a pace of about 5 to 10 words per second;
- keeping key content inside the on-screen safe zone so the interface does not cover it;
- using a native, less polished style featuring real people, and ending with a clear call to action.
Instagram adds a constraint that matters if you plan to post the same script to Reels. In an April 2026 post on rewarding original creators, Instagram says accounts that mainly post unoriginal Reels, photos or carousels will no longer be recommended, and that adding a border, watermark or caption credit to someone else’s video does not make it original. AI scripts are fine; reposting a trending creator’s video with a new caption is not.
What a short-video script needs to contain
A short-video script is not a blog post read aloud. It is a timed plan with three parallel tracks: what is said, what is shown, and what text appears on screen. Not every viewer has the sound on, so on-screen text has to carry the message on its own, while the visuals give people a reason to keep looking.
| Part | Job | Typical length in a 30-45 second video |
|---|---|---|
| Hook | Stop the scroll and state what the viewer gets | One sentence plus one visual, first 1-3 seconds |
| Setup | Show why this matters to this viewer | One or two lines |
| Payoff | Deliver the promised tip, result or answer | Most of the video, in 2-4 beats |
| Close | One action: comment, save, follow, watch part 2 | One line |
The timings above are a planning convention, not platform rules. The point is proportion: the hook is short, the payoff gets the time, and the close asks for one thing.
The four-prompt workflow
Run these in one conversation in Claude, ChatGPT or Gemini so each step can see the previous one. Replace the square brackets. The workflow separates the creative decision (which hook) from the production work (shots and timing), because mixing them produces a script built around whatever hook the model wrote first.
Prompt 1: The creative brief
The brief gives the model the facts it cannot guess. Most generic AI scripts come from skipping this step.
You are a short-form video scriptwriter for [PLATFORM: TikTok /
Instagram Reels / both]. Before writing anything, read this brief.
Creator: [WHO IS ON CAMERA, e.g. "a freelance accountant, talking
head, filmed on a phone in her office"]
Audience: [WHO, WHAT THEY ALREADY KNOW, e.g. "first-year freelancers
who have never filed quarterly taxes"]
The one point of this video: [ONE SENTENCE]
What the audience usually gets wrong about it: [YOUR REAL
OBSERVATION]
Proof I can show on camera: [SCREEN, OBJECT, RESULT, DOCUMENT]
Target length: [SECONDS]
Goal of the video: [comments / follows / saves / clicks to profile]
Words and claims to avoid: [e.g. "guaranteed", income claims,
medical claims]
Reply with three questions you would need answered to write a strong
script, then stop.Why each field exists: “the one point” prevents scripts that try to cover five tips in 30 seconds. “What the audience gets wrong” is the raw material for a hook that surprises, and it has to come from you, because the model only knows the most common version of every topic. “Proof I can show” anchors the video in something visual. The final instruction to ask questions and stop catches gaps before the model fills them with assumptions. If you want to understand why the opening role line helps and where it does not, see our guide to role prompting.
Prompt 2: Hooks built on different mechanisms
Asking for “10 viral hooks” returns ten versions of the same sentence. Asking for hooks that each use a named mechanism produces real variety you can choose from or test.
Write 6 hooks for this video, each using a different mechanism:
1. Contradiction: challenge something the audience believes.
2. Specific mistake: name a mistake they are probably making.
3. Result first: show or state the end result, then rewind.
4. Direct question about their situation.
5. Insider detail: something people outside the field rarely see.
6. Visual hook: an action or object on screen, with minimal words.
For each hook give:
- SPOKEN: the first line, max 12 words
- ON-SCREEN TEXT: max 8 words, readable without sound
- FIRST SHOT: what the viewer sees in the first second
Rules:
- Every hook must be delivered on by the rest of the video. No
promises the script cannot keep.
- No invented statistics, earnings or "studies show".
- No clickbait phrases such as "you won't believe" or "nobody is
talking about this".The three-part output matters because a hook is not just a sentence: the first shot and the on-screen text do as much work as the spoken line, especially for sound-off viewers. The rule that the video must deliver on the hook protects watch time, since a viewer who notices the promise will not be kept has every reason to swipe away. The ban on invented statistics is there because models reach for fake numbers when told to be attention-grabbing.
Prompt 3: The timed script with a shot list
Once you pick a hook, the model turns the brief into a production-ready script. The table format makes it easy to film from.
Use hook [NUMBER]. Write the full script as a table with columns:
TIME | SPOKEN LINE | ON-SCREEN TEXT | SHOT / ACTION
Constraints:
- Total length: [SECONDS] seconds. Assume about 2.5 spoken words per
second and flag if the script runs over.
- Change the shot or on-screen element at least every 3-4 seconds.
- On-screen text must make sense with the sound off.
- Keep text away from the bottom and right edge of the frame, where
the app interface sits.
- One call to action at the end, matching the goal: [GOAL].
- Write spoken lines the way [CREATOR] talks: short sentences,
contractions, no corporate words.
After the table, list any claim in the script that I should
double-check before filming.The 2.5 words per second figure is a planning assumption for conversational speech, not a platform number; adjust it after you time yourself reading one script aloud. The instruction to change the visual every few seconds follows TikTok’s advice to use transitions, text and graphics to hold attention. The safe-zone line is generic on purpose: exact safe-zone dimensions vary by placement, so check the current spec in the platform before you design text-heavy frames. The claims list at the end turns the model into a fact-check assistant instead of a fact source.
Prompt 4: Cut and tighten
First drafts of short scripts are almost always too long. A separate editing pass works better than asking for brevity up front.
Edit this script as a ruthless short-form editor.
1. Cut every word that does not change the meaning.
2. Remove any sentence that repeats the hook's promise instead of
delivering on it.
3. Point out the moment where a viewer is most likely to swipe away,
and rewrite it.
4. Give me a version 10 seconds shorter than the current one.
Return both versions as tables in the same format.Asking for the most likely swipe-away moment forces the model to read the script from the viewer’s side. The shorter version gives you a cut-down to test, which is cheap to film at the same time as the main take.
Worked example: from brief to shot list
This example is illustrative, written for this guide. The creator is a freelance bookkeeper; the one point is that freelancers should move a fixed share of every payment into a separate tax account the day it arrives; the common mistake is waiting until the tax deadline to work out what they owe. The goal is comments.
Hooks the workflow might produce:
- Specific mistake: “If you only think about tax when it’s due, this is for you.” On-screen: “Freelancer tax mistake #1”. First shot: a calendar page with the deadline circled.
- Result first: “This boring bank account is why my clients never panic in April.” On-screen: “The account every freelancer needs”. First shot: phone screen showing two bank accounts side by side.
- Visual: no line; the creator drops an invoice into a jar labeled “not mine”. On-screen: “Part of every invoice isn’t yours”.
Script excerpt using the result-first hook (35 seconds):
TIME | SPOKEN | ON-SCREEN TEXT | SHOT
0-3 | This boring bank account is why my | The account every freelancer | Phone: two accounts
| clients never panic at tax time. | needs | side by side
3-8 | Most freelancers spend everything that | Waiting = panic | Creator to camera
| lands, then work out tax at the deadline.| |
8-16 | Instead: the day a payment arrives, move | Every payment -> move a % | Screen recording of
| a fixed percentage into a separate | | a transfer
| account you never touch. | |
16-26 | Your accountant can tell you the right | Ask your accountant for your | Creator writing a %
| percentage for your income and country. | number | on a sticky note
26-32 | Then tax day is a transfer, not a crisis.| Tax day = one transfer | Transfer confirmed
32-35 | Comment "account" and I'll share how I | Comment "account" | Creator points down
| name mine so I don't spend it. | |Notice what the script does not do. It does not give a specific percentage, because the right number depends on the viewer’s income and country, and a confident figure in a 35-second video is the kind of claim that misleads people. That is exactly the line the “claims to double-check” step in prompt 3 should flag. The close asks for one action, and the comment keyword gives the creator something concrete to follow up on.
Before and after: one generic hook
The difference between a generic prompt and the workflow shows most clearly in the hook.
- Prompt: “Write a viral TikTok hook about freelancer taxes.” Typical result: “Stop scrolling! Here are 5 tax tips every freelancer NEEDS to know!” It promises a list the video may not deliver, says nothing specific, and sounds like thousands of other videos.
- Workflow result: “If you only think about tax when it’s due, this is for you.” It names a situation the viewer recognizes, implies a single fix and matches the creator’s calm, practical tone.
Mistakes that make AI scripts sound like AI
Most problems with AI-written short-video scripts come from what the prompt left out, not from the model.
- No observation from you. Without “what the audience gets wrong”, the model writes the most common take on the topic. Your experience is the only part a viewer cannot get elsewhere.
- Hook and payoff mismatch. A hook promising a secret followed by common advice loses viewers. Prompt 2’s delivery rule exists for this.
- Written for reading, not speaking. Long sentences and formal words sound stiff on camera. Read the script aloud before filming and cut anything you stumble on.
- Everything said, nothing shown. If the shot column is “talking head” for every row, the video relies entirely on audio. Add at least one visual proof.
- Too many calls to action. “Like, comment, follow and share” asks for four things and gets none. Pick the one that matches the goal.
- Fake numbers and claims. Models invent statistics and earnings figures when asked to sound exciting. For finance, health or legal topics, check every claim and avoid specific advice that depends on personal circumstances.
- Copying a trend without adding anything. Using a trend format is normal; reposting someone else’s video is what Instagram’s originality rules target.
To test hook variations systematically rather than one by one, the same principles from our guide to A/B testing ad copy with AI prompts apply: change one element at a time and judge by retention and the action you asked for, not by views alone.
Making it a repeatable process
Once the workflow produces scripts you like, save the creator description, audience, tone rules and banned words as standing instructions in a Claude Project, a custom GPT or a Gemini Gem. Each new video then only needs the one point, the common mistake, the proof and the length. Pasting two or three of your best past scripts as examples helps the model match your voice; our explainer on zero-shot vs. few-shot prompting covers how to choose and format those examples.
For volume, plan topics first and script second. Our prompts for planning a month of social media content produce a topic list you can feed into prompt 1 one video at a time. If you already have long-form content, repurposing one blog post into multiple pieces is a reliable source of single-point video ideas, since each section of a good article is often one short video.
Finally, review results by hook mechanism, not just by video. After a few weeks of posting, you will see whether contradiction, result-first or visual hooks hold your particular audience better, and you can tell prompt 2 to weight toward the mechanisms that work for you.
