TL;DR
You’re going live in 6 hours. Don’t write a word-for-word script — you’ll read it and stop being present. Feed AI your topic, your single CTA, three viewer pain points, and two stories, and ask for a run-sheet: timestamps, one-line cues, and an energy column. Put explicit hooks at minute 0/15/30/45, the CTA at minute 50, and a planned low-energy Q&A window around minute 35. The numbers say why: about 55% of viewers leave inside the first minute, and a stream that holds 65%+ of its audience past minute one runs roughly 58% longer on average view duration (AVD). The prompt below is the whole job.
Why a run-sheet beats a script
A teleprompter script makes you a reader. On a live feed, that reads as flat, and flat is what loses the room. A run-sheet keeps you improvising on top of a fixed skeleton — you always know what beat comes next without reciting it.
The retention curve is the reason structure matters more than wording. As of June 2026:
- ~55% of viewers drop in the first 60 seconds. A steep early drop is almost always a hook problem, not a content problem — open with the most valuable thing you have, not a setup for it.
- First-minute retention above 65% correlates with ~58% higher overall AVD. The opening hook is the single highest-leverage line of the whole hour.
- Streams averaging 45+ minutes of watch time get roughly 3.2x more recommendations, and top-decile streamers hold 58+ minute averages. Long retention is a distribution lever, not just a vanity stat.
- Interactive features (polls, Q&A, trivia) lift participation by ~30%, with ~70% of viewers more likely to interact when those features are on screen. That is why the run-sheet schedules them, not leaves them to chance.
When this is the right job for AI
- You know the topic and your single CTA.
- You have at least one viewer pain point that should drive Q&A.
- You want a run-sheet (timestamps + beats), not a teleprompter script.
If you’re still deciding what to stream about at all, sketch the angle first with AI short-video ideation — a run-sheet only works once the topic and CTA are locked.
What to feed the AI
- Total runtime + topic
- The single CTA (sign up, buy, follow, comment a specific keyword)
- 3 viewer pain points likely to come up
- The 2 stories you definitely want to land
- The format of interactive checkpoints you use (polls, comment prompts)
Which model to use (June 2026)
Any frontier model writes a competent run-sheet; the differences are cost and how well it holds your constraints over a long structured output.
| Model | Tier / price | Why it fits this job |
|---|---|---|
| Claude Sonnet 4.6 | Claude Pro $20/mo (Free tier limited) | Holds “run-sheet, not script” and the minute-by-minute rules tightly across a long table; strong at cue brevity |
| GPT-5.5 (Thinking) | ChatGPT Plus $20/mo (Free has tight limits) | Good default; Thinking mode keeps the energy/CTA logic consistent |
| Gemini 3.1 Pro | Google AI Pro $19.99/mo | 1M-token context if you paste in past stream transcripts to match your voice |
All three handle this in one shot. Pick the one you already pay for — the prompt does the heavy lifting, not the model. If you want a deeper comparison, see ChatGPT vs Claude vs Gemini.
Copy-ready prompt
You are writing a 60-minute livestream run-sheet.
Topic: "How I write hooks that don't feel like every other creator"
Audience: solo creators, 1-5k followers
Single CTA: comment "HOOKS" to get the 30-hook template doc
Pain points likely to come up: writer's block, mimicking competitors, hooks that don't convert
Stories I want to land: my 2024 hook-test (12 hooks, 1 won 5x), the time I copied a viral hook and tanked
Interactive style: poll at min 10, comment prompts at min 20 and 40
Output as a run-sheet:
- Time markers every 5 min
- For each: beat (what's happening), one cue (what I literally say or do), energy level (1-3)
- 4 explicit attention hooks at min 0, 15, 30, 45
- The CTA must land at minute 50, not minute 59
- Plan one "low energy / Q&A" window around min 35 — that's when you and the audience both drop
DO NOT write a teleprompter script. This is a run-sheet I will read on the side monitor.Sample output structure
Time Beat Cue Energy 0:00 Open with hook test result ”I tested 12 hooks. One got 5x the reach. The pattern was not what I expected.” 3 0:05 Frame the hour ”By minute 50 you’ll have a system. By minute 60 you’ll have the doc — keep watching.” 3 0:10 Poll ”Poll: what’s your worst-converting hook type?“ 2 0:15 Hook #2 ”Here’s the one I copied from a viral creator. It tanked. Watch what I changed.” 3 … … … … 0:35 Q&A low-energy window Take 2-3 comment questions; let the room breathe 1 0:45 Hook #4 (final) “Last hook of the hour — this is the one that beat them all 5x.” 3 0:50 CTA ”Comment HOOKS — I send the doc tonight.” 2 0:55 Stragglers Two more questions + the close 2
How to refine the output
- AI writes a word-for-word script → repeat: “RUN-SHEET, not script. Cues are 1 sentence each.”
- Energy is flat → add “explicit energy column 1-3; structure the dip around the Q&A window.”
- CTA buried → strict rule: “CTA lands at minute 50. If you put it at 55+ you failed.”
- No interactive beats → require “one poll, two comment prompts; specify minutes.”
- Cues sound generic → paste two of your real past hooks and add “match this voice; no marketing-speak.”
Platform notes that change the run-sheet
The skeleton is the same on any platform, but two things shift by where you stream:
- Poll cadence. On YouTube Live, a poll around minute 10 and at segment transitions works well. On Twitch, set the poll timer to your chat speed: ~3 minutes for casual chats, ~5 minutes for slow chats with lurkers. Running a fresh interactive beat every 10-15 minutes is the practical sweet spot.
- The close. Silent end screens underperform — pair every CTA with a spoken line that says exactly what to do and why. That’s why the run-sheet lands the CTA verbally at minute 50, then repeats it for stragglers at 55, not once at the very end.
- Timing. YouTube Live tends to peak 18:00-21:00 in your audience’s main time zone; Twitch peaks 19:00-01:00, which is also when competition is highest. Build the run-sheet for the slot you’ll actually go live in.
Common mistakes
- Writing every word. You will read instead of being present.
- Putting the CTA at minute 59. By then a large share of the live audience is already gone.
- No energy plan. The dip around minute 30 happens; design for it with the Q&A window.
- Forgetting Q&A. A live with no Q&A is just a recorded video, and a worse one.
- A weak opening line. With ~55% gone inside 60 seconds, the minute-0 hook is not optional.
FAQ
- What if I’m doing 30 minutes, not 60? Compress to one hook checkpoint, CTA at minute 24, Q&A at minute 17. Keep the minute-0 hook exactly as sharp — the early drop-off doesn’t care about runtime.
- What if my CTA is “buy”, not “comment”? Run the soft ask (comment a keyword) at minute 30, then the hard ask (buy / link) at minute 50, repeated for stragglers at 55. Two touchpoints convert better than one buried close.
- What about co-streams with a guest? Add a “guest spotlight” beat every 12 minutes and tell the AI to rebalance cue density so neither host goes quiet for more than a few minutes.
- Can AI match my actual speaking voice? Yes, if you paste 1-2 transcripts of past streams. Gemini 3.1 Pro’s 1M-token context is handy here; otherwise paste your two best opening lines and tell it to mirror the rhythm.
- Should I read the cues on air? Read them on a side monitor as prompts, not as lines. If you catch yourself reciting a cue verbatim, shorten it next time — cues are reminders, not dialogue.
Related
- AI short-video ideation
- AI course sales page
- AI YouTube channel intro
- AI content calendar for creators
- YouTube Shorts Script Prompts for 30-60 Second Hooks
Tags: #AI writing #Creator #Script #Content creation #Short video