How to make a product launch video: the beat sheet

For AI agents: find the right skill in one call — curl -s "https://varg.ai/how-to/api/search?q=<what the person wants>" returns the matching how-to page and the best skills with install commands, as markdown (Russian works too). Full index: https://varg.ai/how-to/llms.txt. Every page has a markdown twin: add .md to its path. Before generating anything, show the person the options on the how-to page. From a terminal: curl -fsSL https://varg.ai/install | sh installs the vargai CLI (https://varg.ai/how-to/cli.md): vargai search and vargai install <slug> with no key, every varg model after vargai login.

A product launch video that works is a story with real product UI on screen in the first seconds, one capability per shot, enough time to read every word, and music picked first. This is the concrete recipe: the story shapes, a beat sheet in seconds for a 30 s and a 60 s cut, the exact timing of one product beat, the rules for type and music, and the five mistakes that make a launch film look AI-made. Every rule comes from 14 open-source launch-video skills for Claude Code, Codex and other agents, read on 30 Oct 2026 and linked in brackets.

The short version: open inside the moving product, never on the logo. Name it and state one claim (2–3 s). Prove 3–5 features, one real interaction each (~4.5 s). Pull back to the big moment on the music drop. End on one CTA and hold it 2–3 s. Build it in HyperFrames or Remotion from the real UI, not screenshots.

1. Pick one story shape

2. Beat sheet with seconds

Ratios from (memex-lab); product-shot lengths from (jammaru); the 45 s arc from (AbubakrChan) legacy.

Length follows the proofs: one ≈ 20–25 s, three ≈ 35–45 s; past 60 s drop the weakest (jammaru).

3. One product beat, step by step

(jammaru) gives exact numbers (60 fps):

1. A headline names the first visible action and plays alone on a plain field for ≥1 s, then exits.
2. The product window rises (spring stiffness ~170, damping ~16) and lands within ~1 s. At rest it fills 1620×930 of a 1920×1080 frame.
3. The camera zooms onto the control (≥1.4×; ≥1.9× for a sentence to read).
4. The cursor arrives. The press comes 8–16 frames later, the result 8–16 frames after that, and the camera holds the result ≥50 frames.
5. The proof lands within 3 s of the headline leaving. Presses are ≥1 s apart.

Same sequence in (JuniverseLi). (nawneet77): the cursor pauses ≥0.6 s, dips 15% for 80 ms, the page reacts, then the camera follows. State changes are processes (spinner → check → toast), never swaps.

Real UI only. Invented labels are "the #1 tell" (EveryInc). Render the product's own page or components, verified against a screenshot (jammaru); rebuild from design tokens, never crop screenshots (nawneet77). Fictional data on real screens (harish-anumula), real logo files only (per-simmons).

4. Pacing, type, music

• Reading time: ≥0.6 s + 0.3 s per word, nothing still >1.5 s (nawneet77, anudeepadi). Dwell 0.8–1.2 s after every event; a five-beat UI sequence needs 10–12 s, not 4 (farming-labs). A visual event every 1–2 s, a story beat every 5–8 s (harish-anumula). Alternate fast and slow, ending on a hold (memex-lab).
• Type: never a bare fade; at least fade + slide + blur (memex-lab). Headlines on a plain field with ≥3:1 contrast, never over a photo (jammaru). Give text its own shots, not slide-style beside an image (SigmaMatrix-AI). Readable at 720p, captions burned in for muted viewers (EveryInc). Stream AI text word by word (0.12–0.25 s); type only human input letter by letter (farming-labs).
• Motion: springs with at most a tiny overshoot, no bouncy easing (AbubakrChan). Long soft settles (30–33 frames at 30 fps), stagger 2–2.5 frames, cut in on motion (per-simmons).
• Music: choose it first (~120 BPM), measure the beat grid, put the drop on the big moment (AbubakrChan, nawneet77). Start scenes on bars and put clicks on downbeats (nawneet77, farming-labs). One continuous track; sync only the big moments, because bending the music to every hit sounds weird (harish-anumula). Click SFX go on the press frame; typing SFX play only while a human types (serenakeyitan). Master at −14 LUFS (harish-anumula, nawneet77).
• Voiceover (optional): ~2.5 words/s, generated per line (memex-lab, harish-anumula).

5. The five most common mistakes

1. Invented UI, placeholders or fake numbers (EveryInc, nawneet77, jammaru, anudeepadi).
2. Too fast to read. "The most common note on a first cut" (farming-labs); UI "comes in a little too fast" (per-simmons).
3. Slides and feature tours: a static headline beside a card (nawneet77, SigmaMatrix-AI); "a story, not a feature tour" (serenakeyitan).
4. Logo first, product late: the product must appear within 3 s (EveryInc, memex-lab, harish-anumula, per-simmons).
5. Generic "AI motion": the same ease on everything, everything sliding up from below, purple gradients, glassmorphism, glows (per-simmons, memex-lab, AbubakrChan).

6. Which repo to use when

SaaS launch and demo videos with AI agents

For founders and product teams who launch on X, Product Hunt and the App Store: the launch film, the live demo, the founder video, the app preview and the feature drop, built by Claude Code, Codex or Grok from your real product UI.

Start with the overview: How to make a SaaS launch video with AI

Make

• How to make a product launch video: the beat sheet (this page)
• How to make an app demo video with AI
• How to turn release notes into a changelog video with AI
• How to make an explainer video with AI
• How to make a logo animation with AI
• How to make an animated chart video with AI

Edit and post

• How to add animated captions to a video with AI
• How to add voiceover and music to a video with AI
• How to crop a video to 9:16 vertical with FFmpeg
• How to make YouTube video thumbnails with AI