Most AI YouTube scripts sound the same: “In today’s video, we’re diving deep into…” followed by a list of facts and a long outro. Viewers click away in seconds. The fix is not a better one-off prompt, it is a reusable process that bakes in what keeps people watching: a hook with no greeting, a clear plan, a pull into every next section, and a fast payoff. In this guide I share a free Claude Skill, youtube-script-writer, that does exactly that for long videos and Shorts, with a timed table of voiceover, on-screen text and B-roll you can hand straight to an editor or a text-to-speech voice.
- What it is: a
SKILL.mdthat turns your topic, notes and target length into a retention-focused script, laid out as Time | Voiceover | On-screen text | Visual. - Length maths: about 150 spoken words per minute, so an 8-minute video is roughly 1,200 words and a 60-second Short about 130-160.
- Built-in honesty: any fact you did not provide gets a [verify] mark, so you do not publish made-up numbers.
- Works in: claude.ai (upload as ZIP), Claude Code (
~/.claude/skills/) and the API. Pair it with the hook writer and SEO pack from the YouTube skills pillar.
Why a skill beats an “AI script generator” prompt
A typical script prompt is one sentence long: “write a YouTube script about X”. The model fills the gaps with its defaults, which are a polite welcome, generic structure and a summary at the end. Every one of those defaults costs you viewers.
A Claude Skill is a saved set of instructions that Claude loads automatically when your request matches its description. Because it is loaded every time, the rules you care about are always applied: no greeting before the hook, one idea per section, a re-hook every minute or so, [verify] marks on facts, and your channel’s tone. If you want the background on how skills load and where they work, read the Claude Skills guide.
Anatomy of a script that holds attention
| Part | Time (long video) | Job | Common mistake |
|---|---|---|---|
| Hook | 0:00-0:15 | Show the result, name a mistake, or ask the viewer’s exact question | “Hey guys, welcome back to the channel” |
| Proof and plan | 0:15-0:45 | Why trust this, and the 3-5 parts coming | Long backstory before any value |
| Body sections | Most of the video | One idea each, mini-promise at the start, pull line at the end | Lists of facts with no reason to keep watching |
| Re-hooks | Every 60-90 s | A question, example or visual reveal that resets attention | Same pace and visual for minutes |
| Payoff | Near the end | Deliver exactly what the title promised | Payoff hidden or missing |
| Close | Under 20 s | Point to one specific next video | Long outro, “like and subscribe” monologue |
Shorts compress this: one hook line, three to five quick beats, the payoff, and a last line that loops back to the start so a replay feels natural. YouTube Shorts can now be up to 3 minutes long, but most clips work best well under a minute.
The youtube-script-writer skill
Copy the file below or use the download button, and save it as youtube-script-writer/SKILL.md. The YAML header follows Anthropic’s rules (lowercase hyphenated name under 64 characters, description under 1,024 characters, no reserved words).
---
name: youtube-script-writer
description: Writes retention-focused YouTube scripts (long videos or Shorts) with a hook, open loops, timed sections, B-roll and on-screen text cues, ready for voiceover or text-to-speech. Use when the user asks for a YouTube script, a voiceover script, a faceless video script, or to rewrite or tighten an existing script.
---
# YouTube script writer
## Inputs (ask only for what is missing)
- Topic and working title
- Target length (minutes) or "Short"
- Who the viewer is and what they should be able to do afterwards
- Facts, steps, notes or sources the script must use
- Face on camera, voiceover, or text-to-speech voice
Use the channel brief tone if one is available.
## Length math
Spoken English runs about 140-160 words per minute. Target words = minutes x 150. A 60-second Short is roughly 130-160 words. Tell the user the word count at the end.
## Structure (long video)
1. Hook (0:00-0:15): result, mistake, question, contrast or stakes. No greetings.
2. Proof and plan (0:15-0:45): why trust this, and the 3-5 parts coming up.
3. Body sections: one idea per section. Each section opens with a mini-promise and ends with a line that pulls into the next ("but that only works if...").
4. Re-hook every 60-90 seconds with a pattern change: a question, a quick example, a visual reveal.
5. Payoff: deliver the promised result clearly.
6. Close (under 20 seconds): one specific next video to watch and why. No long outro.
## Structure (Short)
Hook (1 line) > 3-5 fast beats > payoff > a loop line that connects back to the start.
## Output format
A table with columns: Time | Voiceover | On-screen text | Visual/B-roll.
After the table: word count, estimated length, and a list of every fact that needs checking, marked [verify].
## Writing rules
- Write for the ear: short sentences, contractions, one idea per sentence.
- For text-to-speech, spell out numbers and abbreviations the way they should be read, and avoid symbols in the voiceover column.
- Never invent statistics, quotes, studies or prices. Use only facts the user provided or facts from a source you can cite; mark anything else [verify].
- The script must add original value (explanation, commentary, testing or story). Do not rewrite someone else's video or article line by line. Save it inside a folder named youtube-script-writer
Install and use it
claude.ai
- Settings > Capabilities: turn on code execution and file creation (“Create and edit files”).
- Zip the
youtube-script-writerfolder (the ZIP must contain the folder, not just the file). - Customize > Skills > + > Create skill > Upload a skill, then choose the ZIP and make sure it is toggled on.
Claude Code
mkdir -p ~/.claude/skills/youtube-script-writer
# save SKILL.md there, then in Claude Code:
/youtube-script-writer "Why your GPU runs out of VRAM" 1-minute ShortPrompts that work well
- “Write an 8-minute script about setting up Ollama on Windows for beginners. My notes are below. Faceless, my own voiceover.”
- “Turn this blog post I wrote into a 5-minute script. Keep my opinions, cut the filler.”
- “Tighten this script to 6 minutes and fix the hook. Mark anything you changed.”
- “Write a 45-second Short version of section 3.”
Best results come when you paste real notes: steps you tested, your opinion, numbers you trust. The skill builds the structure; your notes provide the substance that makes the video yours.
Example output: a 60-second Short
This is the format the skill returns, using a topic from this site. I wrote the notes it was based on (VRAM rules of thumb from our VRAM guide), so the facts are not invented.
| Time | Voiceover | On-screen text | Visual |
|---|---|---|---|
| 0:00 | Your graphics card isn’t too slow for AI. It’s too small. | NOT SLOW. SMALL. | Task Manager, VRAM bar at 100 percent |
| 0:04 | Every AI model has to fit in video memory, called VRAM, or it crawls. | VRAM = the limit | Simple box filling up animation |
| 0:10 | Rule one: a model in four-bit needs roughly half a gigabyte per billion parameters, plus room for context. | ~0.5 GB per 1B (4-bit) | Calculator overlay |
| 0:20 | So an eight billion model fits on an eight gig card. A seventy billion model doesn’t. | 8B fits / 70B doesn’t | Two cards side by side |
| 0:29 | Rule two: longer chats use more memory, so cut context length if you hit errors. | Shorter context = less VRAM | Settings slider in LM Studio |
| 0:38 | Rule three: if it still doesn’t fit, use a smaller quant like Q4 instead of Q8. | Q8 → Q4 | File list showing sizes |
| 0:47 | Do those three things and your “slow” GPU suddenly isn’t. | 3 rules. Done. | Chat streaming fast |
| 0:53 | Which brings us back to your card: it was never slow. | (loop) | Back to first shot |
Word count: about 120, roughly 50-55 seconds at a natural pace. Nothing needed a [verify] mark because every number came from the notes.
Example outline: an 8-minute video
For long videos the skill writes the full table. Here is the outline level so you can see the shape it aims for (target about 1,200 words):
| Time | Section | Pull line into the next part |
|---|---|---|
| 0:00 | Hook: the finished local chatbot answering a question offline | “Here’s how to get this in ten minutes, even on a laptop.” |
| 0:15 | Plan: install, first model, pick the right size, fix the two common errors | “But first, one setting most guides skip.” |
| 0:45 | Install Ollama, check the version | “Now the fun part, but don’t download the biggest model.” |
| 2:30 | Choosing a model size for your GPU (re-hook: “guess which one my laptop could run”) | “If you see this error, don’t panic.” |
| 4:30 | Two common errors and fixes | “Once it runs, make it actually useful.” |
| 6:00 | Use it in a nicer chat app | “Here’s the result, side by side with ChatGPT.” |
| 7:20 | Payoff comparison | – |
| 7:45 | Close: one next video (best models for your GPU) | – |
Making scripts sound good with text-to-speech
Faceless channels often use a synthetic voice. Scripts written for reading sound robotic when spoken, so the skill writes for the ear. A few extra tips:
- Spell numbers as words where the reading matters: “eight gig card”, “twenty twenty-six”. “8GB” may be read as “eight G B”.
- Remove symbols from the voiceover column (%, /, &, →). Keep them in the on-screen text column.
- Use punctuation for pacing. Commas and full stops are the most reliable way to control pauses across voices.
- Test pronunciations of model names, brands and acronyms once, then add the spelling that works to your channel brief.
- Re-record the hook yourself if you can. Even one human line at the start helps.
For voices, my ElevenLabs alternatives guide compares cloud and free local options, and the Fish Audio Drama 3 local guide shows a free expressive TTS model on your own GPU.
Claude skill vs dedicated script tools
| Option | Strengths | Weaknesses | Best for |
|---|---|---|---|
| Claude + this skill | Fully editable rules, your tone, works with your own notes and files, can chain with SEO and thumbnail skills | No built-in YouTube data; you supply research or enable web search | Creators who want control and consistency |
| Plain chatbot prompt | Zero setup | Generic structure, rules forgotten between chats | One-off scripts |
| YouTube-specific AI tools (for example vidIQ’s AI features) | Channel data, keyword ideas and trends inside one app | Less control over structure and voice; subscription costs | Idea and keyword research alongside scripting |
| Writing by hand | Most original | Slowest | Signature videos and personal stories |
Many creators combine them: research and keyword ideas in a YouTube tool, then structure and drafting in Claude with the skill, then a personal edit pass.
Fixing scripts with your retention graph
After you publish, YouTube Studio shows an audience retention graph for each video (Analytics > the video > Engagement). Use it to improve the next script:
| Graph shape | What it means | Script fix |
|---|---|---|
| Big drop in first 30 seconds | Hook did not match the title/thumbnail, or started too slowly | Run the hook writer, start with the result |
| Steady slide all the way | Pacing is flat | Add re-hooks every 60-90 s and cut repeated points |
| Sharp drop mid-video | A boring section or a broken promise | Shorten or move that section, add a pull line before it |
| Spikes (rewatches) | Viewers rewatch something valuable or confusing | Make more of it, or explain it more clearly |
| Drop at the outro | Normal | Keep the close under 20 s and point to one video |
Paste what you see into Claude with the script and ask it to revise the weak section. The youtube-analytics-review skill in the YouTube skills pillar can also rank your videos from a Studio CSV export.
Troubleshooting
| Problem | Fix |
|---|---|
| Script still starts with a greeting | The skill may not have triggered. Say “use the youtube-script-writer skill” or check the toggle under Customize > Skills. |
| Script is too long or too short | Give a target in minutes; the skill aims for about 150 words per minute and reports the count. Ask it to trim to an exact number of words. |
| Too many [verify] marks | Paste your notes or sources. With web search enabled, ask it to check and cite each marked fact. |
| Sounds generic | Add 3 example scripts or transcripts you like to the Project knowledge, plus your channel brief. |
| Upload error on claude.ai | Zip the folder, not the file; folder name must be exactly youtube-script-writer. |
| Table hard to paste into an editor | Ask for “voiceover only, one paragraph per line” as a second output for your teleprompter or TTS tool. |
FAQ
Is it OK to use AI-written scripts on a monetised channel?
Yes. YouTube’s policies target mass-produced, repetitive or reused content, not AI assistance. A script built on your notes, research and opinions, read with care and matched to real visuals, is original content. A script that rewrites someone else’s video is not.
How many words is a 10-minute YouTube script?
About 1,400 to 1,600 words at a natural speaking pace of 140-160 words per minute. Tutorials with lots of on-screen demos can use fewer words because the visuals carry time.
Can the skill write in languages other than English?
Yes. Ask for the script in your language. Speaking speed differs between languages, so check the length by reading the hook aloud and adjusting.
Can it write scripts for YouTube Shorts and Reels?
Yes, it has a Short structure (hook, quick beats, payoff, loop line). For turning an existing long video into several Shorts, use the Shorts repurposer skill.
Does it work with Claude’s free plan?
You can paste the skill body into a chat as instructions on any plan. Uploading it as a custom skill on claude.ai is listed for Pro, Max, Ultra, Team and Enterprise in Anthropic’s docs; check Customize > Skills on your account. It also works in Claude Code.
Why does the skill avoid “like and subscribe” at the start?
Asking for a subscription before giving value gives viewers a reason to leave. A short, specific call to action after the payoff, or pointing to the next video, keeps sessions going, which helps the channel more.
Final Thoughts
Good scripts are mostly structure plus honest substance. The skill handles the structure every single time, and the [verify] rule keeps the substance honest. Feed it your real notes, read the hook aloud before recording, and use your retention graph to tune the next one. Next steps: write the hooks with the hook writer from the YouTube skills pillar, then package the upload with the YouTube SEO skill.
