Build a private, self-hosted "watch a movie together in sync" site for two people in different countries — synced play/pause/seek over WebSocket, WebRTC camera+mic call with a coturn TURN relay, a shared laser pointer, ffmpeg smooth-bitrate transcoding so it never buffers, and a token-gated room link. Use when someone wants to watch local video files in sync with a partner/friend remotely, replace Kosmi/Teleparty/Twoseven with something they own, stream a personal movie library over the web with synced controls, or add a video-call overlay to a video player. Triggers: "watch movies together long distance", "sync video playback two browsers", "self-hosted watch party", "Kosmi alternative", "stream my movie to my girlfriend with subtitles and sync".
Subject-aware reframing of horizontal video to vertical (9:16 / 4:5 / 1:1) using YOLOv8 person detection. Per-scene decision between TRACK (crop tightly on subject) and LETTERBOX (scale + black bars) so people stay centered through cuts and motion. Trigger when user says "reframe to vertical", "16:9 to 9:16", "convert landscape to portrait", "shorts crop", "smart crop for TikTok", "vertical YouTube short from horizontal", "auto-crop interview clip", "reframe podcast video", "make this video vertical", or any equivalent intent. Use as a STAGE inside viral-clipper / book-video / channel-breakdown pipelines whenever the source has more than one person on screen or the subject moves around the frame.
Receive and verify Clerk webhooks. Use when setting up Clerk webhook handlers, debugging signature verification, or handling user events like user.created, user.updated, session.created, or organization.created.
Automatically extract reusable patterns from Claude Code sessions and save them as learned skills for future use.
Strip AI voice from text — humanize while preserving meaning. Combines glebis's RU/EN/DE/ES/FR/PT/JP/IT diagnostic pipeline with the author's Humanizer Prompt + the Buddha-at-2-AM shipped-manuscript reverse-engineering + the (redacted) Russian-original voice analysis. Use for KDP books, the publication series, marketing copy, or any text that needs to read like a real person wrote it. Pairs with locked voice voice spec.
Invoke DeerFlow — ByteDance's multi-agent research harness running Claude Sonnet 4.6. Use for deep research tasks, multi-step web research, document analysis, and report generation. Spins up a local LangGraph agent that decomposes the task into sub-agents in parallel. Requires the backend to be running first.
Router for all UI/UX/visual design work. Invoke when user types "design conductor", "/design", "design:", or "redesign:". Reads the task, picks the right stages (PLAN → STYLE → BUILD → REFINE → AUDIT), and dispatches the matching sub-skills in order. Don't merge the sub-skills — chain them.
Fetch brand-specific DESIGN.md system specs from getdesign.md when the user wants UI to match a known brand's aesthetic ("make it look like Stripe", "Apple-style", "Linear feel"). Covers 59 brands across SaaS, fintech, consumer tech, automotive, AI labs, and creator tools. Use BEFORE picking a generic style skill — brand match always beats generic.