silent-ugc
Clone viral TikTok UGC app-ad hooks with your own AI actor and app demo. Use when the user wants to recreate a viral TikTok ad, build a TikTok account around one consistent AI creator, mass-produce UGC ad variations, or swap an actor into an existing hook. Runs on the Arcads MCP (nano-banana-2, gpt-image-2, Seedance 2.5).
jasonlee-breadcrumb · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add jasonlee-breadcrumb/silent-ugc --skill silent-ugc技能说明
name: silent-ugc description: Clone viral TikTok UGC app-ad hooks with your own AI actor and app demo. Use when the user wants to recreate a viral TikTok ad, build a TikTok account around one consistent AI creator, mass-produce UGC ad variations, or swap an actor into an existing hook. Runs on the Arcads MCP (nano-banana-2, gpt-image-2, Seedance 2.5).
Silent UGC: TikTok hook cloning and AI actor accounts
Take a viral TikTok app ad and turn it into as many variations as you want. Keep the proven hook, put your own AI actor and app demo in it, and rotate the on-screen text. Everything runs on the Arcads MCP (nano-banana-2, gpt-image-2, and Seedance 2.5).
The format you are recreating: a person reacting to something off screen, a line of text over their face, then a hard cut to a screen recording of the app. The hook is generated fresh for each variation. The app demo is one real recording reused across all of them.
Two modes, ask which one at the start
Clone mode recreates one video. The user gives you a TikTok URL. Run the interview, then the pipeline once.
Account mode builds a TikTok account of 10 to 20 videos around one AI actor. The user gives you a TikTok video, an actor image, the app, and how many videos they want. Every scene reuses the same face in a new room, outfit, and light, so a scroller reads the whole account as one person. Get the user to approve the full scene and hook list, plus the cost, before you generate anything.
Phase 1: interview
The app. What is it? A store link is enough. Fetch it and read the category, features, pricing, audience, and reviews. This is what drives the hook angles. Ask what the goal is: installs, trials, or awareness.
The two assets you need. Search the working folder first, and only ask if they are missing.
- App demo video. A screen recording of the app in use (look for mp4 or mov files named demo, screen, app, or recording). Strip the audio. If it has a hook attached, split it and keep only the demo.
- Actor image. The face for every video (jpg or png named model, actor, face, or reference). Frontal, well lit, at least 700px. If it is low resolution or a stock photo the user may not have rights to, say so. A generated face avoids both problems.
The source TikTok. The viral video whose hook you are cloning.
Phase 2: verify the Arcads MCP
Call arcads_list_products. If it responds, note the productId and continue. If the tools are not available:
To connect the Arcads MCP:
- Get your MCP connection details from the Arcads dashboard (Settings, then API / Integrations / MCP).
- In Claude Code:
claude mcp add arcads <MCP URL>, then restart the session.- In Claude desktop or web: Settings, Connectors, add the Arcads connector and authorize. Tell me when it is added and I will check again.
This skill handles the source video on its own. yt-dlp downloads it, ffmpeg pulls the frames, and the Read tool renders those frames directly, so you do not need a separate video-watching skill. The hooks are silent, so there is no transcript to fetch either.
Phase 3: pick the hook and suggest the text
Propose 10 to 20 on-screen texts, grouped by angle, strongest first:
- Price against the human alternative. "my trainer charges $80 for what this app does free"
- Confession. "i wasted 2 years before finding this"
- Gatekeeping. "trainers are mad this app exists"
- Mistake callout. "you're doing X wrong and don't even know it"
- Beginner anxiety, matched to the app's audience.
- Lazy. "i stopped planning X, the app does it now"
- Transformation. "day 1 vs day 60, i'm shook"
Keep each under 12 words, casual lowercase, with a curiosity gap that pairs with a reaction shot. Re-texting a video you already generated costs nothing, so that is the cheapest way to scale a rotation.
Phase 4: confirm before you run
Show the plan and get an explicit yes:
- How many videos, which hook goes on which scene, and which source each scene is built from.
- The actor you are using.
- The cost (see the cost model at the end).
Do not start generating until the user confirms.
The pipeline, per video
1. Download and split
TikTok's JS challenge fails now and then, so retry instead of stopping on the first error:
for i in 1 2 3 4 5; do yt-dlp -o "video.%(ext)s" "<TIKTOK_URL>" && break; sleep 3; done
Given a profile URL instead of a single video, list it first and pick the ad, or ask:
yt-dlp --flat-playlist --print "%(id)s | %(view_count)s views | %(title).70s" --playlist-end 20 "<PROFILE_URL>"
Find the boundary between the hook and the demo:
ffmpeg -i v.mp4 -vf "select='gt(scene,0.3)',metadata=print" -an -f null -
Take the first big scene score, then confirm it by eye. Cut the hook, mute it, and re-encode it for frame accuracy:
ffmpeg -i v.mp4 -t <cut> -an -c:v libx264 -crf 18 -preset fast hook.mp4
Keep the hook. The user's own demo replaces the source's demo.
2. Build the reference image with nano-banana
This is your actor, in the scene, holding the pose. It becomes the reference image for Seedance.
For a new scene, generate the actor directly in a fresh room with arcads_generate_image_nano_banana (model nano-banana-2, 9:16). Pass the actor image as the reference and describe the room, outfit, light, and gesture. Keep the same face and hair every time.
To match a specific source scene closely, recreate the source frame first, then put your actor in it:
- Recreate and strip the text (
arcads_generate_image_nano_banana,nano-banana-2, 9:16):
Recreate the first image exactly: same vertical selfie framing, same camera angle, same pose and hand position, same clothing, same [lighting], same background, same amateur iPhone front-camera quality with mild sensor noise and soft compression. Remove all text overlays so the image is completely clean of text. Keep real skin texture, uneven lighting, no beautification, no retouching.
- Swap in your actor (`arcads_ge