YouTube Video Script Writer for Gemini
Write a complete YouTube script with a hook that earns the first 30 seconds, a chaptered body with on-screen cues, and an ending designed for retention, sized to your video length and audience.
Fill in the variables
Prompt
<role>
You are a YouTube scriptwriter who has written for educational and how-to channels that measure success by average view duration, not views. You write for the ear, in short spoken sentences, and you plan what is on screen while each line is said.
</role>
<context>
Video topic: {{topic}}
Channel and audience: {{channel}}
Target length: {{length_minutes}} minutes (about 140 spoken words per minute)
The one thing the viewer should be able to do or decide after watching: {{outcome}}
My angle or opinion (what makes this different from the top results): {{angle}}
Call to action at the end: {{cta}}
</context>
<constraints>
- Spoken language: contractions, short sentences, no sentence a person would not say out loud. Read it aloud in your head.
- The hook (first 20–30 seconds) must state the payoff and the stakes; no "hey guys, welcome back", no channel intro before the hook.
- Chapters every 60–120 seconds, each with a title a viewer would click in the timeline.
- Every chapter includes [ON SCREEN: …] cues for b-roll, text or graphics, and at least one concrete example, number or demonstration.
- One re-hook halfway through: a line that reminds the viewer why the second half matters.
- The ending: a 2-sentence summary, the CTA once, and an "end screen" line that points to the next logical video. No "smash that like button".
- No invented statistics. If a figure is needed and I did not provide it, write [stat needed].
</constraints>
<goals>
- Keep the viewer past 30 seconds and past the midpoint.
- Deliver the promised outcome clearly enough that the comments say "this actually helped".
- Give me a script I can read from a teleprompter without editing.
</goals>
<instructions>
1. Write the title (under 60 characters) and three alternative titles.
2. Write the script in sections: HOOK, then numbered CHAPTERS with timestamps, then the ENDING. Put on-screen cues in brackets on their own lines.
3. After the script, add a 2-line description for the video and 8 chapter timestamps formatted for the YouTube description.
</instructions>
<output_format>
Plain text. Section headings in capitals. Timestamps as mm:ss. Keep the total word count within 10% of the target length.
</output_format>
Example input
Topic: How to batch-edit 100 product photos with an AI background remover Channel: small-business tools channel; viewers run online shops, not design studios Length: 6 minutes Outcome: process a folder of product photos in under an hour without Photoshop Angle: Most tutorials show one photo. I show a real batch, including the 10% that fail and how to fix them fast. CTA: Download the free checklist linked in the description
Example output generated with Claude Sonnet 4.6 from the example input above
TITLE: I Batch-Edited 100 Product Photos With AI (The 10% That Failed)
Alternatives: "100 Product Photos, One Hour, No Photoshop" / "AI Background Removal at Scale: What Breaks" / "Stop Editing Product Photos One at a Time"
HOOK (00:00–00:28)
A hundred product photos. One hour. No Photoshop. That's what I'm doing today, on camera, with a real folder from a real shop.
[ON SCREEN: folder view, 100 thumbnails, timer starts]
Here's why you should stay: ninety of these will come out perfect in about ten minutes. The other ten are the ones every tutorial skips, and fixing them is the whole difference between a listing that looks professional and one that looks like a hobby. I'll show you both.
CHAPTER 1 — Set up the batch (00:28–01:40)
Before you touch any tool, sort the folder. Put reflective items, anything with fur or hair, and anything transparent into their own subfolder. Those are your likely failures and you want to know where they are.
[ON SCREEN: three subfolders being created: "easy", "shiny", "fuzzy"]
Then export everything at the same size. I use 2000 pixels on the long edge. Mixed sizes are the number one reason a batch comes back inconsistent.
[ON SCREEN: export dialog, 2000 px highlighted]
Now open the background remover and look for a bulk upload or a folder option. If the tool only takes one image at a time, it's the wrong tool for this job, however good the single result looks.
CHAPTER 2 — Run the easy ninety (01:40–02:50)
Drag the "easy" folder in. Mugs, boxes, bottles, books. Watch the timer.
[ON SCREEN: upload progress, then a grid of results]
Spot-check ten, not all ninety. Zoom to 200% on edges and shadows. If those ten are clean, trust the batch. Then download as PNG with transparency, because you'll want to drop these on a white or coloured background later and a baked-in white background locks you in.
[ON SCREEN: edge zoom at 200%, clean]
That was eleven minutes for ninety images. [stat needed if you want a per-image cost here]
RE-HOOK (02:50–03:00)
Now the part nobody shows you. Ten images, and three different ways they go wrong.
CHAPTER 3 — The shiny ones (03:00–04:00)
Reflective products confuse the model because the background is literally in the product. Watch what happens to this chrome kettle.
[ON SCREEN: kettle result, reflection half-removed]
The fix is not to fight the tool. Re-shoot or re-crop so the reflection shows a plain surface, or use the tool's "keep" brush to paint the reflection back in. Thirty seconds each. Don't try to fix it in the prompt; there is no prompt for this.
CHAPTER 4 — Fur, hair and lace (04:00–04:55)
Fuzzy edges get either shaved off or haloed. Most tools have an "edge refine" or "hair" mode; it's slower, so you only run it on this subfolder.
[ON SCREEN: before/after of a knitted hat, halo removed]
If your tool has no hair mode, this is the one case where a $5 manual edit per image is cheaper than your time.
CHAPTER 5 — Transparent and white-on-white (04:55–05:30)
Glass and white products on white sweeps come back with holes. The trick is to shoot them on a mid-grey background in the first place; the model needs contrast, not a white sweep. Add that to your shooting checklist and this failure disappears next batch.
[ON SCREEN: glass vase on grey, clean cutout]
ENDING (05:30–06:00)
So: sort first, run the easy ninety fast, and give the hard ten their own ten minutes. One hour, one hundred photos, no Photoshop. The checklist with the folder structure and the shooting rules is linked in the description. If you want to see what I do with these cutouts next, the video on generating lifestyle backgrounds is on screen now.
[ON SCREEN: end screen with "AI lifestyle backgrounds" video]
Description: I batch-edit 100 real product photos with an AI background remover and show the 10% that fail — reflective, fuzzy and transparent items — and how to fix each in under a minute. Free checklist below.
00:00 The challenge · 00:28 Sort and export the batch · 01:40 Run the easy ninety · 02:50 Where it breaks · 03:00 Reflective products · 04:00 Fur, hair and lace · 04:55 Glass and white-on-white · 05:30 Checklist and next video
Why this works & how to customize
Why it works: retention on YouTube is decided in the first 30 seconds and again at the midpoint, so the prompt puts the payoff and stakes in the hook and forces a re-hook. The word budget (about 140 words per minute) keeps the script honest about length; most scripts run long because nobody counted. On-screen cues on their own lines make the script usable as a shot list.
How to customize: for talking-head commentary without b-roll, change the cue rule to "[TEXT: …] lower-thirds only". For Shorts, set the length to 1 minute and delete the chapter requirement. If you have a signature intro, add "insert my 5-second intro after the hook, not before". Feed the model your actual numbers in the angle field so it never needs the [stat needed] placeholder.
How to customize: for talking-head commentary without b-roll, change the cue rule to "[TEXT: …] lower-thirds only". For Shorts, set the length to 1 minute and delete the chapter requirement. If you have a signature intro, add "insert my 5-second intro after the hook, not before". Feed the model your actual numbers in the angle field so it never needs the [stat needed] placeholder.
Comments (0)
Sign in to comment.
No comments yet.