Prompts for YouTube Shorts and Reels That Work
In short
A prompt returns a usable short-form script when it pins down six things: who the viewer is, how many seconds of speech you want, what kind of hook to open on, the single idea the video delivers, the exact call to action, and the output format. Leave any of them unspecified and the model fills the gap with its own average — which is why "write me a viral YouTube Shorts script" produces the same beige listicle every time, for everyone. Four prompts cover the whole pipeline: a hook generator that returns variants you choose from, a script writer that takes one hook plus a word budget, a shot-list prompt that turns lines into things you can actually point a camera at, and a metadata prompt for title, description and pinned comment. Keep them as files in a repo so you can diff them when quality drops. And know the ceiling: YouTube's monetization policy explicitly targets mass-produced, templated content, so the prompt output is a first draft, not a video.
Why "write me a viral script" returns mush
An LLM asked for a short-form script with no constraints resolves every open variable toward the median of its training data: a beginner audience, a 250-word body, a rhetorical-question opener, three tips crammed into thirty seconds, "like and subscribe" at the end. That output is not wrong. It is generic, which on a recommendation feed is the same thing. The fix is not a longer prompt but one that closes the six variables the model would otherwise guess at.
| Variable | What the model does if you omit it | What to write instead |
|---|---|---|
| Audience | Writes for "beginners" — the widest, blandest reader | "backend devs who already use a coding agent daily" |
| Duration | Produces 250+ words, roughly 90 seconds of speech | "35 seconds spoken — 85 words maximum, hard limit" |
| Hook mechanism | Opens with a rhetorical question, every time | "open on a claim most of the audience will disagree with" |
| The one idea | Packs three tips into a thirty-second slot | "one idea only: secrets leak into agent transcripts" |
| Call to action | "Like and subscribe" | "CTA: comment the word diff and I'll send the script" |
| Output format | Prose with headings you then strip by hand | "return JSON: {hook, lines[], broll[], cta}" |
The word budget is the variable people skip and the one that decides whether the script is filmable. Record thirty seconds of yourself reading at your normal pace and count the words instead of trusting a generic words-per-minute figure. Most people land between 2 and 3 words per second, and the gap between those two is a third of your runtime.
The four prompts
These are written as fill-in templates. Angle brackets are slots you replace; everything else stays.
1. Hooks — generate many, pick one.
You write hooks for short vertical video.
Topic: <one sentence>
Audience: <who they are and what they already know>
Angles I will not use: <fake urgency / hype / anything I can't back up>
Return 10 hooks. Each is one spoken sentence, 12 words max, sayable
in under 3 seconds. Vary the mechanism across the set: 2 contrarian
claims, 2 specific numbers, 2 common mistakes, 2 questions the viewer
cannot answer from memory, 2 open loops.
Numbered list, mechanism in brackets after each. No emoji.
Generating ten and choosing one is the point. A single hook is the model's median guess; ten lets you apply the judgement it does not have.
2. Script — one hook in, one filmable script out.
Write the spoken script for one vertical short.
Hook (use verbatim as line 1): <paste the hook you picked>
Length: <N> seconds spoken = <M> words. Do not exceed <M>.
Audience: <who>
After watching, the viewer should be able to: <one concrete thing>
Constraints:
- No greeting, no "in this video". Line 1 is the hook.
- Every sentence under 14 words, sayable in one breath.
- Include this detail, which only someone who has done it would
know: <your detail>
- Final line is exactly this CTA: <your CTA>
Return plain lines, one sentence per line. No timestamps, no
section headers, no stage directions.
The "detail only someone who has done it would know" slot is what keeps the script from sounding like every other script in your niche. If you cannot fill it, you do not have a video yet — you have a topic.
3. Shot list — turn lines into visuals.
Turn this script into a shot list for a <faceless screen-capture>
short.
<paste script>
For each line return: line number, what is on screen, on-screen text
(6 words max), and whether it needs b-roll or a screen recording.
Markdown table.
Where you cannot suggest a specific visual, write TODO. Do not
invent a generic stock shot to fill the row.
That last instruction matters more than it looks. Left alone, models assign "person typing on laptop" to every line, and you find out in the edit.
4. Metadata — title, description, pinned comment.
From this script, write publishing metadata.
<paste script>
Return:
- 5 title options, 60 characters max, each containing a phrase a
viewer would actually type into search
- description: two sentences describing what is shown, plus 3 tags
- a pinned comment asking the one question the script deliberately
left open
Do not promise anything the script does not deliver.
For more starting points, our prompt library has ready-made cards for content work, and the faceless video guide covers the production side these prompts feed into.
Make the output machine-readable
If you live in a terminal, stop pasting these into a chat window. Put each prompt in a file with a placeholder and pipe topics through it — claude -p and most agent CLIs read stdin and write stdout like any Unix tool:
while read -r topic; do
sed "s|<TOPIC>|$topic|" prompts/hooks.md \
| claude -p > "out/${topic// /-}.hooks.md"
done < topics.txt
Two things this buys you. The prompts live in git, so when quality drops after a model update you diff against the version that worked instead of trying to remember what you typed last month. And asking for JSON instead of prose (return {"hook": ..., "lines": [...], "broll": [...]}) lets the output feed a teleprompter, a subtitle file or a batch renderer with no parsing step.
The caveat: batching drafts is fine, batching uploads is what gets channels demonetized. See the last section.
Adapting the templates to your niche
Three edits do most of the work.
Write a voice.md with five sentences you actually said on camera and five phrases you never want to hear back ("in today's video", "without further ado", "smash that like button"), and prepend it to every script prompt. It is the cheapest fix for the "all my scripts sound the same" problem, because you are handing the model a target distribution instead of asking it to be original.
Fill the detail slot from your own work — a build time, an error message, a bill, a thing that broke. That is the part no other channel in your niche can copy.
Then adjust the budget per platform, because the container differs even when the file does not:
| YouTube Shorts | Instagram Reels | |
|---|---|---|
| Maximum length | 3 minutes (YouTube Help) | 3 minutes (Buffer's roundup of the Jan 2025 change) |
| Frame | Vertical; 9:16 is what YouTube recommends for Shorts creatives (Google Ads Help) | Vertical, 9:16 |
| Feed playback quirk | For Shorts ads, "only the first 60 seconds will play on the Shorts feed" | — |
| Practical script budget | 30–60 s for anything algorithmic; longer only if the payoff needs it | Same |
Two notes. The 60-second feed cap is documented for ad creatives, not organic Shorts — do not repeat it as a rule for your uploads. And three minutes is the ceiling, not a target; a 45-second idea stretched to three minutes only widens the window in which people swipe away.
What the prompt will not fix
Retention. The model has no access to your drop-off curve. It cannot tell you that viewers leave at second 7 because the hook promises a payoff you deliver at second 20. Only your analytics can, which means posting, reading the graph, and editing the prompt — not just the next script.
Policy risk. This is the part the "50 viral prompts" posts leave out. YouTube's channel monetization policies require that content "should be your original creation" and "not be mass-produced, generic, repetitive, or manipulative", and call out "AI-generated content made with generic or unoriginal templates giving the impression of mass production without adding the creator's original, authentic insights or perspective" (YouTube Help). The July 2025 rename from "repetitious" to "inauthentic content" was a clarification, not a new rule. Thirty interchangeable scripts is a monetization problem wearing a productivity costume: draft with the templates, then make each video materially different or do not post it.
Delivery and edit. Pacing, cuts, captions, audio levels. A good script read badly is still a bad video.
Trends and audio. Model knowledge is stale by definition. Anything about a current sound or format comes from your own feed, not the prompt.
Your tooling budget. None of this needs a dedicated "AI script generator" subscription. As of July 2026, Claude Pro is $20/month billed monthly, and any general-purpose model on a paid consumer plan runs these four prompts fine. If a tool's only feature is a hidden version of the prompts above, you are renting placeholder text. Our prompting guide covers the underlying technique if you would rather build your own set.
FAQ
What is the best prompt for YouTube Shorts?
There is no single best prompt, but there is a reliable shape: specify the audience, the exact number of spoken words, the hook mechanism, the one idea, the exact CTA, and the output format. The highest-leverage line is the word budget — "35 seconds spoken, 85 words maximum, hard limit" — because it forces the model to cut rather than pad. Generate ten hooks separately from the script and pick one; that split does more than any amount of prompt wording.
How long should a YouTube Shorts script be?
Count words, not minutes. Record yourself reading at your normal pace for thirty seconds and count what you got through — most people land between 2 and 3 words per second. A 30-second Short is therefore roughly 60–90 words including the hook. Shorts can technically run to 3 minutes, but the ceiling is not a target; use the extra time only when the payoff genuinely needs it.
Can I use AI-generated scripts on YouTube and still be monetized?
Yes, with a real constraint. YouTube does not ban AI assistance, but its monetization policies require content to be your original creation and not "mass-produced, generic, repetitive, or manipulative", and specifically name AI-generated content built from generic templates with no original insight. Practically: draft with prompts, then add something only you have — your data, your mistake, your screen recording, your voice. Near-identical uploads from one template are exactly the pattern the policy targets.
Do the same prompts work for Instagram Reels and TikTok?
The structure ports; the numbers do not. Hook, one idea, CTA and shot list are format-agnostic. What changes is the word budget and the CTA that makes sense on each platform — "comment a keyword" behaves very differently on Instagram than on YouTube. Keep one script prompt with length and CTA as slots rather than three near-duplicates that drift apart.
Why do all my AI-written scripts sound the same?
Because you are asking for the median and getting it. Two fixes: give the model a voice sample — five sentences you actually said, plus a banned-phrase list — prepended to every run, and require one concrete detail per script that could only come from your own work. Without those, every prompt in your niche converges on the same output, because it is the same model reaching for the same average.