Viral Replicator + Review-to-Ad

Two paths, same destination: a Higgsfield-ready video prompt.

• Path A — Viral replication. Paste a viral video URL → deconstruct the structure → rebuild it for the user's brand.
• Path B — Review-to-ad. Point at a G2 product reviews URL → cluster reviews into testimonial themes → generate ad scripts.

When to use

User prompts that should trigger this skill:

• "Deconstruct this Instagram video and rebuild it for my brand: <URL>"
• "What made this viral video work? Now make me a version of it."
• "Scrape Castos's G2 reviews and turn them into a testimonial ad."
• "Pull these reviews and write me 3 ad scripts."

Routing

Route on the first user input:

Path A — Viral replication

Step 1 — Fetch the post

node skills/viral-replicator/scripts/fetch-ig-post.mjs <URL> --out output/viral-replicator/<slug>-<YYYY-MM-DD>/raw-post.json

The script captures: caption, hashtags, on-screen text, video poster, duration, engagement counts. If Instagram blocks the headless browser (common — they aggressively rate-limit), fall back to examples/kallaway-fixture.md and tell the user the demo is using a fixture.

The same script handles TikTok and YouTube Shorts URLs at best-effort; for YouTube Shorts especially, ask the user to paste the transcript if scraping fails.

Step 2 — Deconstruct

Load prompts/deconstruct-viral.md and produce a structured breakdown: hook (0–3s), pattern interrupts, narrative arc, visual style, audio role, payoff, CTA shape, why-it-worked hypothesis.

Write to output/viral-replicator/<slug>-<YYYY-MM-DD>/deconstruction.md.

Step 3 — Rebuild for the user's brand

Ask once: "What's the brand and topic this rebuild is for? One sentence." (Skip if the user already told you.)

Load prompts/rebuild-viral.md and produce a shot-by-shot script + a paste-ready Higgsfield video prompt using templates/higgsfield-rebuild-prompt.md.

Write to output/viral-replicator/<slug>-<YYYY-MM-DD>/rebuild.md and higgsfield.md.

Step 4 — Render (optional)

If the Higgsfield MCP is connected, call ToolSearch with query higgsfield to find the current video-generation tool, then call it with the prompt from templates/higgsfield-rebuild-prompt.md. If not, the prompt sits paste-ready.

Path B — Review-to-ad

Step 1 — Scrape reviews

node skills/viral-replicator/scripts/scrape-g2-reviews.mjs <URL> --top 25 --out output/viral-replicator/<brand>-reviews-<YYYY-MM-DD>/raw-reviews.json

The script captures per review: reviewer name (or "Verified User"), role, company size, star rating, "what do you like best", "what do you dislike", quote excerpt. G2 has anti-bot protections — if the scrape is blocked, fall back to examples/castos-g2-fixture.md and tell the user.

For Trustpilot URLs, the script auto-switches selectors. For arbitrary review text the user pastes, skip Step 1 and feed the text directly into Step 2.

Step 2 — Cluster into testimonial themes

Load prompts/review-to-ad.md. Produce 2–3 testimonial themes. Each theme:

• Theme name (the customer truth, NOT a feature name)
• 1 representative pull-quote per theme (verbatim, with reviewer attribution)
• Why this theme converts

Write to output/viral-replicator/<brand>-reviews-<YYYY-MM-DD>/themes.md.

Step 3 — Ad-ify

For each theme, write a 30-second testimonial-style spot:

• Hook (0–3s) — uses the pull-quote or a paraphrase
• Setup (3–10s)
• Demonstration (10–22s)
• Payoff + CTA (22–30s)

Choose the visual treatment per theme:

• Talking-head reconstruction — for emotional / "this saved my business" testimonials
• Text-on-b-roll with VO — for technical / specific-feature testimonials
• Side-by-side before/after — for migration / switching testimonials

Use templates/testimonial-ad-prompt.md to write the Higgsfield video prompt for each spot.

Write to output/viral-replicator/<brand>-reviews-<YYYY-MM-DD>/scripts.md and higgsfield.md.

Step 4 — Render (optional)

Same MCP discovery pattern as Path A.

Output structure

output/viral-replicator/
├── <viral-post-slug>-<YYYY-MM-DD>/
│   ├── raw-post.json
│   ├── deconstruction.md
│   ├── rebuild.md
│   └── higgsfield.md
└── <brand>-reviews-<YYYY-MM-DD>/
    ├── raw-reviews.json
    ├── themes.md
    ├── scripts.md
    └── higgsfield.md

Voice & tone

• For rebuilds: match the structure of the original, NOT the voice. The voice should sound like the user's brand.
• For testimonial ads: keep customer quotes verbatim where possible. Resist polishing — real language reads as real.
• No "transformative." No "game-changing." No "I never thought I could…" unless the customer literally said it.

Demo commands

Deconstruct https://www.instagram.com/p/DAi1lWIPFuB/ and rebuild it for an insulated cup brand.

Scrape https://www.g2.com/products/castos/reviews and turn them into 3 testimonial ad scripts.