--- name: recreate-concept description: Recreate a winning organic-dropshipping video from an Instagram, Facebook, or TikTok link, break down why it went viral, then rebuild it frame by frame with the user's OWN character and product using AI (Nano Banana Pro images + Seedance motion via Higgsfield). Use when the user says /recreate-concept, pastes a viral IG/FB/TikTok link and wants their own version, "recreate this concept", "make this with my product", "rebuild this video for my character", or hands over a proven concept to turn into their own clips. Built on the ANOMALY AI-organic system. Confirm-before-generate is enforced. --- # /recreate-concept, the ANOMALY concept-recreation engine **Job:** take one proven viral video (Instagram / Facebook / TikTok link) and turn it into the user's OWN version, same winning structure, their character and product, generated with AI, so a beginner can ship a viral-shaped video without filming anything. **The system this runs on:** ANOMALY AI-organic dropshipping. Every product gets a character who appears to make/own the product; viewers perceive it as real; the character and the story drive the purchase. We recreate proven concepts because copying what already keeps people rewatching beats inventing from scratch. Ground yourself first (read if present): `mentorship-launch/skool/lesson-docs/claude-project/03-making-viral-videos.md`, `06-winning-database.md`, `07-prompt-library.md`. --- ## THE 2 LAWS (never break) 1. **CONFIRM BEFORE YOU GENERATE.** Show ALL the prompts (image + motion) in a numbered list and WAIT for the user's explicit go before generating a single image or clip. They scan, they fix, then you run. Never generate blind. This is the most important rule in the whole system. 2. **THEIR CHARACTER, THEIR PRODUCT, NOT A COPY.** Keep the winning structure (hook, beats, pacing, why it rewatches). Swap in the user's character and product. Change enough that the algorithm sees it as fresh (fresh character, angle, or variant), never a 1:1 clone. --- ## INPUTS TO COLLECT (ask only what's missing, one at a time) - The **winning video link** (IG / FB / TikTok). - The user's **product** (what it is, price, handmade / wow-factor / problem-solving). - The user's **character**, an image if they have one; if not, offer to research + generate it first (see the character step in `03-making-viral-videos.md`). - A **product photo** if they have one (used as a generation reference). Don't force a long brief. If the character or product isn't ready, say so and offer to do that step first. --- ## WORKFLOW ### 1. Pull the video - Use `yt-dlp` to download the video + metadata (views, upload date, caption). Example: `yt-dlp -o "%(id)s.%(ext)s" ""` and `yt-dlp --dump-json ""` for metadata. - For deep on-screen-style analysis, optionally run `python3 mentorship-launch/brain/tools/analyze-video.py ""` (Gemini watches the frames). Use only when you need frame-by-frame visual detail; otherwise the download + your own read is faster. ### 2. Break down WHY it won Produce a tight breakdown (this is the same as the "Break Down A Winning Video" prompt): - **The hook:** the exact first frame and the on-screen text, and why it stops the scroll. - **The structure, beat by beat:** every clip in order. Map it to a shape, handmade = hook → build clips → end result; wow-factor = hook → buildup → reveal; problem-solving = hook → problem → before/after. - **Why it rewatches:** what keeps people watching to the end (that watch time is what the algorithm rewards). - **Keep vs change:** what to copy exactly, and the one thing to change to make it theirs (character, variant, or angle). Show this breakdown to the user and confirm the direction before building prompts. ### 3. Write the prompts (DO NOT GENERATE YET) Go clip by clip, in order. For EACH clip write: - **Image prompt**, Nano Banana Pro, 2k, recreating that frame with the user's character + product. Compose for a 9:16 vertical, iPhone-handheld / natural-light UGC look (not studio-polished, not model-perfect, that kills the organic feel). Reference the user's character image + product photo. - **Motion prompt**, Seedance 2.0 (or Kling) describing exactly how that image animates (what moves, camera, the action from the original beat). Keep it matched to what the original clip did. Present all prompts as one numbered list. Then STOP and ask the user to review and confirm or edit. Per the higgsfield prompt limits: keep each prompt plain-text and under ~1100 chars; pre-upload references and pass their IDs. ### 4. Generate (only after the user says go) - Images: Higgsfield MCP `generate_image` (Nano Banana Pro), passing the pre-uploaded character + product references (`media_upload` → use the returned id). - Poll `job_status`, then download each result. - Animate: feed each approved image to `generate_video` (Seedance/Kling) with its motion prompt; poll + download. - If a clip comes back wrong, fix that ONE prompt and regenerate just it, don't redo the set. ### 5. Hand off to the edit (CapCut) Give the user the assembly plan (they finish in CapCut): - Put the ORIGINAL viral video on the track as a reference for timing. - Drop the generated clips in order. - Get a fresh copy of the song: Shazam the original's audio → download via YouTube-to-mp3 → line it up to match. - Mark the beats, place each clip on a beat so it flows like the original. - Add the text hook (TikTok Sans), tweak wording + emoji slightly for their product. Close by offering the next step: split-test variations once it's posted, or recreate the next concept. --- ## STYLE OF THE INTERACTION - Beginner-friendly, one step at a time, plain language. - Show a useful first draft fast; don't open with a wall of questions. - Always land on the confirm-before-generate checkpoint before any generation. - Real proof only, if something's a guess (e.g. estimated views), say so.