TL;DR
You can turn a blank page into a usable first-draft YouTube script in about 15 minutes if you feed the model a tight brief and a structure-aware prompt. The math that matters: a 7-minute talking-head video at a conversational 150 words per minute (wpm) is roughly 1,050 spoken words. The single number to defend is your 30-second hold — YouTube data shows roughly 55% of viewers leave within the first minute, so the opening line carries more weight than the rest of the script combined. Use Claude Opus 4.7 or Sonnet 4.6 for natural-sounding long-form prose, then rewrite the cold open and outro in your own voice by hand.
The task
You’re shooting a 5-10 minute YouTube video this week. Writing the script by hand eats 2-4 hours, and the part that matters most — the first 30 seconds — is the easiest to get wrong. A good AI workflow lands a first-draft script with a real hook, retention beats, B-roll cues, and a clear payoff in about 15 minutes, leaving you time to actually edit and shoot.
This guide is for solo creators, educators, indie SaaS founders, and anyone using YouTube for top-of-funnel content. It assumes a standard 16:9 long-form video, not Shorts (different beats — linked at the end).
The numbers you’re writing against
Before you prompt anything, know your targets. These are 2026 benchmarks for long-form YouTube, not Shorts:
| Metric | Target (as of June 2026) | Why it matters |
|---|---|---|
| Spoken pace | 130-150 wpm conversational; 120-140 for technical | Sets word count: 7 min × 150 = ~1,050 words |
| 30-second hold | Keep 70%+ to look strong; ~55% of viewers leave in minute one | The algorithm’s first recommendation signal |
| Average percentage viewed | 40-55% for 5-15 min videos; 60%+ under 5 min | YouTube’s main “is this worth surfacing” proxy |
| Retention beat cadence | A pattern interrupt every 60-90 seconds | Flattens the mid-video drop-off (~60s and ~180s) |
Word count is just target_minutes × your_wpm. Pick the wpm for your delivery and content type, then hold the model to that number.
Pick the model before the prompt
The model you choose changes how much you have to rewrite. As of June 2026:
| Model | Best for | Context | Note |
|---|---|---|---|
| Claude Sonnet 4.6 / Opus 4.7 | Natural long-form prose, least editing | 1M tokens | Workhorse for the full draft; Opus for nuance |
| GPT-5.5 | Punchy hooks, tight structure, listy formats | ~320 pages in-app on Plus | Strong cold opens; can sound formulaic at length |
| Gemini 3.1 Pro | Repurposing long inputs (full transcripts) | 1M tokens | Feed it a 60-min podcast transcript and let it cut |
A practical hybrid: draft the body in Claude (Sonnet 4.6 on Pro at $20/mo is enough), then ask GPT-5.5 for five alternate cold opens and steal the best line. If you’re repurposing, paste the whole transcript into Gemini 3.1 Pro and ask for the three sharpest beats first.
When AI is the right tool — and when it isn’t
Use AI when you already know your audience and your one-sentence promise and you need structure plus first-pass language. Models are genuinely good at chunking content into act 1 / 2 / 3, suggesting B-roll, and writing cold opens around a specific example.
It also shines for repurposing: turn a 60-minute podcast into a 7-minute script by feeding the transcript and asking for the three sharpest beats.
Don’t lean on AI alone for:
- On-camera delivery. Never read AI prose verbatim — it sounds like nobody. Rewrite every line in your own cadence.
- Channel-specific texture. AI doesn’t know your running jokes, your POV, or the inside info that makes a video feel singular.
- Sensitive topics. Mental health, finance, and medical content needs a human subject-matter pass before it ships.
What to feed the AI
- Topic and the one specific angle you’re taking
- Audience: who they are, what they already know, what they don’t
- Promise: what the viewer can do or believe after watching
- Length: target minutes and the word count it implies at your wpm
- Two channels whose tone you like, two things that bore you
The “what bores you” list is underused and quietly important — it’s the fastest way to kill generic AI filler before it appears.
The prompt
Replace the bracketed placeholders with your own values. Set the word-count target yourself (minutes times your wpm) rather than asking the model to do the arithmetic — models still drift on inline math.
You are a YouTube scriptwriter. Write a [target_minutes]-minute talking-head script.
Topic: [topic]
Angle: [sharp_angle]
Audience: [who_they_are_and_what_they_know]
Promise: [what_they_can_do_after]
Tone references: [channel_1], [channel_2]
Avoid: [what_bores_you]
Target word count: [minutes x your_wpm, e.g. 7 x 150 = 1050]
Structure:
- 0:00-0:15 cold open: a concrete moment or surprising claim. No "in this video..."
- 0:15-0:30 hook: state the promise plus a tease of the payoff.
- Act 1 (~25% of runtime): set up the problem with a specific example.
- Act 2 (~50%): the meat — 3 sub-points, each with one concrete demo or visual.
- Act 3 (~20%): payoff plus one actionable next step.
- Outro: CTA and what to watch next.
Every 60-90 seconds, add a retention beat (pattern interrupt, callback, or question).
Mark B-roll cues inline as [B-ROLL: ...].
Mark on-screen text as [TEXT: ...].
Hit the target word count within 10%; do not pad.What good output looks like
A timed script with timestamp markers, plus [B-ROLL: ...] and [TEXT: ...] directions inline. Keep paragraphs short (1-3 sentences) so you can read them off a teleprompter without losing your place. Label the CTA and the “next video” suggestion clearly so your editor can find them in one pass.
How to check the output
- Read it aloud with a timer. If you’re off the target runtime by more than ~20%, cut or expand. At 150 wpm, 30 seconds over means roughly 75 words to trim.
- Pressure-test the first 30 seconds. Does the opening line work without context? Would you keep watching at second 15? This is the half of the script that decides whether the algorithm surfaces you.
- Count retention beats. You want roughly one every 60-90 seconds. If they’re all questions or all callbacks, vary the type so the rhythm doesn’t feel mechanical.
Common mistakes
- “In this video, I’m going to tell you…” openings — they bleed the 30-second hold
- No retention beats, so viewers drop at the predictable ~60s and ~180s marks
- Missing B-roll cues, which can triple editing time
- A promise the video never actually delivers (kills average percentage viewed)
- AI prose read verbatim, which sounds robotic on camera
Close the loop with YouTube Studio
After the video is live, open the Audience retention report in YouTube Studio. It labels four moments: the Intro (your 30-second hold), Top moments (where almost no one left), Dips (where viewers skipped or quit), and Spikes (rewatched or shared segments). A steep drop in the first 15-30 seconds followed by a flat line means your hook failed; a gradual even decline means pacing, not the open.
Paste the next script back into the model with a note on where the curve dipped last time and ask it to predict where this one will lose people — then fix those beats before you shoot. See Google’s own Measure key moments for audience retention for how to read each label.
FAQ
Which AI is best for YouTube scripts in 2026? For natural long-form narration that needs the least editing, Claude Sonnet 4.6 (or Opus 4.7 for nuance). For punchy hooks and tight structure, GPT-5.5. For repurposing long transcripts, Gemini 3.1 Pro’s 1M-token context. Many creators draft in Claude and borrow a cold open from GPT-5.5.
How long should a 7-minute script be in words? About 1,050 words at a conversational 150 wpm. Slower technical delivery (120-140 wpm) lands closer to 840-980 words. Always set the number yourself rather than trusting the model to do the math.
Can AI write the thumbnail copy too? Yes — ask for five thumbnail text options and pick the one that contradicts an expectation or promises a concrete payoff. Keep it to 3-4 words so it reads on mobile.
How do I keep my own voice? Rewrite the cold open and outro by hand. Those two sections carry the most personality, and they’re the parts viewers remember.
What about Shorts? Different format, different beats — denser pacing (140-160 wpm) and a hook in the first 1-2 seconds. Use the Shorts prompts linked below.
Related
For 60-second formats, use YouTube Shorts script prompts and the broader short video script with AI workflow. For Reels and TikTok openings, swap in reel hook prompts.