ViralForge

ViralForge is an agent-native studio for viral short videos. Paste a link or a local file and get a Viral DNA report that deconstructs the video script: the hook, shot-by-shot structure, word anchors, beat grid, caption style, emotional curve and conversion logic. Replace the product and script, and ViralForge rebuilds the winning structure in one click as an editable VFML project. Then generate 10 costed variants for A/B testing. Not another black-box MP4.

Keywords: viral video replication · video script deconstruction · one-click video remix · AI agent · MCP server · VFML · FFmpeg · UGC ads · content repurposing

English is the default documentation language. Chinese documentation lives in README.zh-CN.md. The v0.1 studio ships with Chinese-language demo fixtures; the documentation and agent surfaces are English-first.

Watch the demos

Three 13–15s English cuts. Click a poster to play the MP4 in your browser — no sign-in, no platform footage.

Prefer the reasoning? The full DNA breakdowns, replica scripts and variant matrices are in Case studies. Every clip is a self-made, AI-labelled demo; the reference footage is never redistributed.

Why ViralForge

Operators already deconstruct viral videos by hand: pause on the first three seconds, guess the hook, note the shot rhythm, copy the caption, find the BGM, re-shoot with their product. It takes hours, depends on taste, and the conclusion stays in one person's head.

The hard part is not editing. It is not knowing what will work. ViralForge turns that tacit knowledge into a shareable, structured, editable asset — and makes the whole loop callable by an agent.

Built for: growth teams, UGC agencies, e-commerce brands, solo founders, and agent builders who want a reproducible video workflow instead of a prompt lottery.

Case studies

Three worked examples from the demo. Each shows the reference DNA, the replica script and the variant matrix. The references are self-made simulations and the replica films are pre-rendered; metrics are fixture values that demonstrate the report format, not customer results. Costs are CNY demo estimates; the app does not bill.

Case 01 · GLOW LAB: a 15s serum demo, rebuilt for a different product

Reference: Douyin · 15s · 96 BPM · 3.12M views · 187K likes · 41.2% completion · 78% 3s retention

Hook: "Stop paying the stupid tax." A counter-intuitive negation followed by outcome-first proof.

What the DNA says

• The first 0.8 seconds reject the viewer's last purchase, creating a "did I get it wrong?" tension. A skin-scanner close-up immediately turns the claim into visual evidence.
• The product does not appear until 0:05. The first five seconds sell the problem, not the bottle.
• Captions land about 0.2 seconds before the voice, so the viewer feels they know before they hear.
• The ending asks for a sample, not a sale. Lower decision cost, higher comment-section intent.

Transferable formula: negation → 3 concrete pains → one ingredient → 7-day proof → low-friction CTA.

Replica script

View the case dossier · run npm run dev and pick GLOW LAB in the studio.

Case 02 · AER: a 14s tote film, rebuilt around a 14-inch laptop

Reference: Xiaohongshu · 14s · 112 BPM · 1.86M views · 121K likes · 38.6% completion · 74% 3s retention

Hook: "Every bag that fits a 14-inch laptop looks like it's headed to a meeting." Identity contrast plus a specific number.

What the DNA says

• "14-inch laptop" is more credible than "large capacity". The viewer does not have to imagine; they only have to confirm their laptop fits.
• It says the stereotype out loud first — big bags look corporate — then breaks it. The contrast only works because the prejudice is acknowledged.
• Three macro shots carry the premium feel: vegetable-tanned leather, stitching, hardware.
• Three scenes map to three audiences, so the comments split into three camps and the engagement rate climbs.

Transferable formula: identity contrast → 3 scenes → material and pocket details → one bag, three uses → color close.

Replica script