ChatGPT Prompts for High-Retention YouTube Scripts

YouTube is the channel most marketers respect and avoid at the same time. Pew’s 2025 platform research has 84% of US adults using YouTube, more than any other platform it tracks, and a third visiting several times a day. The avoidance comes from the production cost, and the most expensive part is the one nobody budgets for: a script that people stop watching ninety seconds in, which quietly wastes the shoot, the edit and the upload behind it.

Scripting is also the part a model genuinely helps with, because YouTube tells you exactly what a script must do and most scripts ignore it. These prompts are built around the retention mechanics YouTube itself reports, and they follow the same rule as everything in my prompt library: the model structures your material, it never invents your facts.

What YouTube actually measures

Before prompting anything, it pays to know what the scoreboard says. In YouTube’s own audience retention documentation, the first metric is called Intro: “Intro tells you what percentage of your audience still watched your video after the first 30 seconds.” YouTube adds that a strong intro number usually means “the content in the first 30 seconds matched the viewer’s expectation of the video’s thumbnail and title.” The rest of the report is spikes, the moments rewatched or shared, and dips, the moments skipped or abandoned.

I read that documentation as a brief. The first 30 seconds have one job, keeping the promise the thumbnail made, and the body of the script has one job, avoiding the stretches that become dips. Every prompt below exists to serve one of those two jobs.

What to gather before you prompt

My kit for a YouTube script takes fifteen minutes to assemble: the topic and the specific claim or payoff the video delivers, the evidence behind it (your data, your examples, your screen recordings), the planned title and thumbnail concept, one transcript of a past video whose tone you liked, and the retention graph of your last two videos if you have any history at all. The transcript matters most. Written-English scripts read fine and sound like a hostage statement, and one pasted example of how you actually talk fixes more than any instruction, for the reasons covered in the few-shot prompting guide.

The prompts

1. The hook, written against the thumbnail

You are a YouTube scriptwriter. The video's title is
[TITLE] and the thumbnail shows [DESCRIBE]. The payoff
the video actually delivers is: [THE SPECIFIC PAYOFF].

Write five hook options for the first 30 seconds. Each
must: restate the promise of the title in the first
sentence, prove within two sentences that the payoff is
real (using only the evidence pasted below), and say or
show what happens next. No "welcome back", no channel
intro, no "before we start". Conversational spoken
English, 60-80 words each.

Evidence: [PASTE]

The hook prompt takes the title and thumbnail as inputs because that is what the viewer just clicked. YouTube’s documentation says intro retention is about matching that expectation, so a hook written without the thumbnail in front of the model is guessing at the one thing it must not guess at. I ban the greeting rituals explicitly because the model has read a million scripts that open with them, and they are the first thing a dip graph punishes.

2. The retention outline

Using the hook chosen above and the evidence pasted
earlier, build a script outline that schedules the
payoffs. For each section: the claim it makes, the
evidence or demonstration it uses, roughly when it lands
(minute marks), and the open question that carries the
viewer to the next section. The full promised payoff
must be delivered - do not hold it hostage to the end,
and do not pad any section that has made its point.
Flag any section that has a claim but no evidence.

This is pacing as a planning artifact. The instruction that matters is delivering the promised payoff instead of dangling it, because bait-and-delay is the pattern viewers punish hardest and the trick every retention guide quietly recommends. My rule is the opposite: pay early, then keep paying. The flag on evidence-free sections is the same anti-invention fence as everywhere else in this library.

3. The spoken draft

Draft the full script from the outline. Rules: spoken
English as in the sample transcript below, sentences a
person can say in one breath, numbers rounded the way
people talk, every claim traceable to the pasted
evidence. Mark [B-ROLL: ...] where the screen should
show something other than the presenter, and mark
[CHECK] on any line where you were tempted to add a
fact that is not in the evidence.

Sample transcript of my voice: [PASTE]

The [CHECK] marker earned its place the first time I ran this prompt. Models fill gaps with plausible fluency, and a script read aloud to camera is the worst possible place to discover an invented statistic, because re-shooting costs what editing text never does. Every [CHECK] either gets a source or gets cut before the teleprompter sees it, the same discipline as the fact-checking pass on written drafts.

4. The dip audit

Read the draft as a bored viewer with one thumb on the
seek bar. List the five moments most likely to become
retention dips: throat-clearing, repeated points,
setup with no payoff, tangents, or anything a viewer
already got from the title. For each, quote the lines
and propose the cut or the tightening. Do not propose
adding anything.

An adversarial pass, and deliberately cut-only. Nearly every retention problem I have seen in a draft script is a length problem wearing a disguise. Vidyard’s own 2023 viewing data found 65% of viewers finish a business video under a minute long and 20% finish one past twenty minutes, and while your mileage varies by format, the direction never flips: the burden of proof is on every extra minute, not on the cut.

5. The Shorts extraction

From the finished script, extract three candidate
Shorts. Each: one claim, its proof, under 45 seconds
of spoken words, with a first line that works with no
context because the viewer arrived from a feed, not
from a click. Note the timestamp range in the main
video each Short is cut from.

One script becoming four assets is the same extraction pattern as my repurposing workflow, pointed at video. The constraint that Shorts open context-free is the whole prompt. A Short that starts with “so as I was saying” is a long-form clip, not a Short, and it performs like one.

6. Title, description and chapters

For the final script: ten title options under 60
characters that promise only what the video delivers,
a two-paragraph description whose first sentence works
as a search snippet, and chapter markers with names a
viewer would actually click. Nothing in the title or
description may promise more than the script contains.

The honesty constraint is retention strategy, not ethics theater. An overpromising title buys a click and sells it back as a 20-second abandonment, which YouTube’s intro metric then reports as a failing video. I would rather rank a title that fills its own shoes.

7. The retention post-mortem

Here is the retention data from the published video:
the intro percentage, and the timestamps of the three
biggest dips and any spikes. [PASTE] Here is the
script. [PASTE] For each dip, quote the script lines
at that timestamp and diagnose the likely cause. For
each spike, do the same. End with three rules to apply
to the next script, each tied to a specific timestamp
in this one.

This prompt is the reason the workflow improves instead of repeating. YouTube hands you a per-second audit of your writing, which is a luxury no blog post ever gets, and most channels never read it against the script. Mine is the last prompt I run and the first input to the next video’s hook.

Why YouTube earns a marketer’s script time now

There is a second audience for these scripts beyond subscribers. When Ahrefs studied 75,000 brands to find what correlates with visibility in AI assistants, branded YouTube mentions came out on top at 0.737, ahead of every content-volume metric they tested. I would not build a channel for that correlation alone, but it stacks the payoff: the same video that holds human attention becomes part of what the machines cite, which is the logic of the zero-click strategy applied to a camera.

FAQ

Can ChatGPT write a whole YouTube script by itself? It can produce something script-shaped in one prompt, and the result reads like everyone else’s one-prompt script. The seven-step version exists because hooks, pacing and evidence are separate problems, and a model handles each well only when it is asked one at a time. The chain costs maybe twenty minutes over the single prompt, which is nothing against a shoot day.

How long should a marketing video on YouTube be? As long as the evidence and demonstrations genuinely fill, and my honest reading of the completion data says that is shorter than most drafts. Run prompt 4 and take the cuts. A tight eight minutes beats a padded fourteen every time I have tested it.

Do these prompts work for faceless or voiceover channels? Yes, and the B-roll markers in prompt 3 become the main output instead of a side note. The hook and dip logic is identical because retention does not care whether a face is on screen.

Should I write the script or the thumbnail first? Thumbnail and title first, always. Prompt 1 cannot do its job without them, and YouTube’s own documentation frames intro retention as expectation-matching. A script written before its promise is a speech looking for an occasion.

This set is part of my prompt and workflow library. The voice setup it assumes is in the brand voice prompts, the extraction pattern comes from the repurposing workflow, and the wider practice starts at the complete guide to AI marketing.