Demo video script creator

Write the script a video is built on: what the voice says, how it says it, and what is on screen while it says it. The output is two files, a narration script the user (or the ElevenLabs API) turns into a voiceover, and a shot plan that demo-video-creator builds from.

The voiceover becomes the video's clock, so the script decides the runtime, the pacing and every scene. Get it right here and the build is mechanical.

Ground rules

• Read the subject before writing a line. Clone or open the project, read its README, docs or chapters, and run or open it. Every claim, command, output and number in the script must be one you saw. If a line can't be shown on screen or checked in the source, cut it.
• Never invent facts, quotes, stats or outcomes. Real-world framing is good; made-up incidents, lawsuits, company names or percentages are not. Prefer "projects get this wrong all the time" over a specific story you can't source.
• Runtime: 3 to 5 minutes by default. Go longer only when the topic needs the depth to be understood, and say so up front with the target length and why. Go shorter only when asked (a teaser, a Short).
• Size the script by words, not feel. Jeremy on ElevenLabs v3 has run between 2.2 and 2.8 spoken words per second with this format (2.23 on one script, 2.77 on another: shorter lines and fewer ... read faster). Plan at 2.5: 450 words ≈ 3:00, 625 ≈ 4:10, 750 ≈ 5:00, and say the range ("about 3:45 to 4:45") rather than one number. Count spoken words (tags excluded) before handing over. Once the .mp3 arrives, measure the real pace and tell the user the actual runtime.
• Write for the ear, in the ElevenLabs format. Follow references/elevenlabs-format.md: one short line per beat, delivery tags at each mood change, ... for pauses, acronyms and file names spelled the way they should be spoken.
• Clean, clear, professional, hands-on. Plain words, one idea per line, no hype words ("revolutionary", "game-changing"), no filler. The viewer should finish knowing one idea well and how to try it themselves.
• Honest about limits. If the subject has a disclaimer (not legal advice, read-only, experimental), it goes in the script.
• Don't record or publish anything. This skill writes files. Generating the voiceover and building the video belong to demo-video-creator; uploading belongs to the user.

Pick the format

Pick the format from what the user asked for and the subject. Don't ask unless it's genuinely unclear; a workshop or lab repo means Explainer + demo.

Workflow

1. Gather. Read the subject end to end: README, docs or chapters, the site, and real command output (run it). Note the real names, commands, outputs and links the script will use.
2. Find the hook. One concrete moment the viewer recognises, framed as a question or a pain ("You find the perfect library on GitHub… were you allowed to?"). It comes first, in the first ten seconds.
3. Outline the acts for the chosen format, with a word budget per act that adds up to the target runtime.
4. Write the narration in the ElevenLabs format. Use first person ("I built") when the user built the thing. End on a short closing line said slowly, ideally echoing the title.
5. Write the shot plan: a table of acts with the script lines each one covers, what's on screen, and whether it's an animated scene or live capture. Note the brand/theme source, the links for the end card, and the audio intent (music bed per act, transition SFX, a click on every real click).
6. Check it. Count spoken words against the runtime; read every factual line back against the source; confirm every live scene shows something that really exists.
7. Hand off. Save inbox/narration.txt and plan.md in a video work folder, give the user the script inline (ready to paste into ElevenLabs), the word count and expected runtime, then continue with demo-video-creator. It gets the voiceover: through the API when the key can use the voice, otherwise by asking the user for the .mp3.

Examples