Video & Audio Production Skills

Turn a dialogue, a story with dialogue, or a one-line idea into finished multi-character audio - a radio drama, a two-host podcast, or clean lip-sync voice clips - 100% locally with no cloud or API keys. Orchestrates three local sibling skills: voice-clone-narration (one cloned/designed voice per character, a clean clip per line), sound-effects (foley/ambience via Stable Audio Open Small), and bg-music (instrumental score/beds via ACE-Step), then mixes them with ffmpeg using sidechain ducking (music ducks under voices+SFX and swells in the gaps) and optional stereo spatialization (pan + distance + movement). Three modes - theater (dramatized play + SFX + music), podcast (two hosts + music bed), and lipsync (clean per-line clips + a manifest for the talking-head skill). When there is music it delivers the full mix plus music-only and no-music stems, plus a timecoded markdown transcript. Use when the user asks for an audio drama, audioteatro / radioteatro, a dramatized dialogue, a podcast episode from a script or idea, a voiceover for lip-sync, or a local TTS conversation with sound.

Generate royalty-free instrumental background music tracks (10-600s, typically 30-120s for reels) as MP3 from a text prompt (genre, mood, tempo, instruments), 100% locally with no cloud or API keys. Uses ACE-Step 1.5 (MIT-licensed weights, MLX-native on Apple Silicon) for commercial-safe music generation, and can mix a generated track under a narration voiceover with automatic ducking to finish reel audio. Use when the user asks for background music, a bg track / BGM for a reel or demo video, royalty-free or original music, local text-to-music generation, an instrumental loop, or to add music under a voiceover.

Generate a brand logo and a matching, brand-consistent set of visual assets - logo marks, wordmark lockups, monograms, app icons, favicons, social avatars, banners, seamless patterns, spot illustrations, and UI icons. Local-first: it prefers on-device generation via the image-gen skill (FLUX.2 Klein on Apple Silicon) and only falls back to a CLOUD image API (Google Gemini / Nano Banana Pro, or OpenRouter) when local can't run - no API key is bundled, it is auto-discovered from your environment or another installed skill and cached outside the repo. Crisp on-brand text is set from real fonts with wordmark.py (no misspelled diffusion text), --look presets steer the visual finish, and a set is kept cohesive via a shared palette and, on the cloud path, the chosen logo as a reference. Use when the user asks to create a logo, brand identity, brand kit, visual identity, brand style, wordmark, app icon, favicon, or a consistent set of on-brand graphics/assets.

Inspect and edit Microsoft Word .docx files while preserving formatting, styles, and layout - 100% locally with python-docx, no cloud or API keys. A small CLI (docxtool.py) lists every document element with a stable index (paragraphs, tables, images, headers/footers), then applies surgical edits: run-aware find/replace that keeps bold/italic/font/color even across split runs, insert paragraphs with a chosen style, edit or append table cells, add table rows copying the previous row's formatting, delete elements, and list available styles. Use when the user wants to read, inspect, edit, modify, update, or fill in a Word document or .docx file; do find-and-replace in a .docx while keeping formatting; change or add table rows/cells; insert headings or paragraphs; work with headers/footers; or template a report, contract, or letter - without opening Word.

Teach the agent how a website works by recording a real user-driven Chrome session, then replay the learned action with different inputs. Phase 1 launches Chrome via Playwright codegen while the user performs an action (e.g. "check flight availability from X to Y on date Z"), capturing all network activity as a HAR plus the UI steps. Phase 2 distills the HAR into a reusable lesson (endpoints, payloads, parameter knobs, auth surface) and reports what actions were learned. Phase 3 performs a variation of the same action for a similar task — either by replaying the learned API calls with substituted parameters (e.g. a different date or destination) or by replaying the UI navigation and recording an mp4 as proof. Phase 4 (optional) turns a recording into a standalone, shareable skill (Cursor personal/project or skills.sh format) that performs one inferred action with the session (HAR) embedded, so it works independently of this skill or the original recording. Use when the user asks to teach/learn a website action, record or capture a browsing session, save network activity to a HAR, "learn how this site does X", replay an action with different parameters, produce UI mp4 proof of a web flow, or generate/build a reusable skill from a recorded session.

Build and use an external, persistent "mental workspace" (a J-space) for a coding agent, 100% locally with no cloud or API keys. Analyze a topic (via the Perplexity MCP or web search) or a codebase (static scan) into a weighted knowledge graph, compile it into a matrix of concepts with local embeddings (all-MiniLM-L6-v2), and derive a thought-sequence keyword ordering. Then load that workspace into active memory, query it with spreading activation to see which concepts light up, edit it, checkpoint it, and even "hypnotize" the agent - pin concepts, install post-hypnotic triggers, and write an always-on induction rule so the workspace primes the agent every session. Use when the user asks to build a mental workspace / knowledge graph / concept map for a topic or codebase, construct or load a J-space, think in more advanced terms about a subject, get thought-sequence keywords, prime/hypnotize the agent toward a set of concepts, or persist and checkpoint a conceptual workspace.

Turn an mp4 video of a single object (orbited 360°) into both a clean, browser-navigable 3D Gaussian splat and a watertight, 3D-printable STL/GLB mesh - 100% locally on Apple Silicon, no cloud or API keys. Shares the video-to-splat pipeline (frame extraction with ffmpeg, camera-pose Structure-from-Motion with pycolmap, Gaussian-splat training with Brush on Metal/WebGPU) and adds two object-specific stages: automatic splat cleanup (isolate the object from background, support surface and floaters via opacity/scale filtering, RANSAC plane removal and DBSCAN clustering) and mesh extraction (Poisson reconstruction to a watertight, millimeter-scaled STL for slicers plus a colored GLB for web/QuickLook). Use when the user wants to scan / photogrammetry / 3D-capture a physical object from a video and get a printable 3D model or a cleaned object splat, mentions turning a video into an STL / 3D print / GLB / mesh, "object to 3D", "video to STL", or cleaning a Gaussian splat of an object.

Read and create PDF documents entirely on-device - no cloud, no API keys. For reading, uses Docling (IBM's document AI) to turn a PDF into structured JSON, Markdown, text, or HTML with real layout understanding: extracted tables (as headers/rows + HTML), section/heading hierarchy, page images, and optional OCR for scanned pages (native macOS Vision OCR). For writing, uses fpdf2 to render a professional, multi-page PDF from a simple JSON config - titles, headings, justified paragraphs, styled tables with alternating rows, images with captions, bullet/numbered lists, key-value KPI blocks, dividers, headers and footers with page numbers. Use when the user wants to read, parse, extract, or inspect a PDF (text, tables, or structure), scrape data out of a PDF, OCR a scanned PDF, or generate/create a polished PDF report, invoice, proposal, or data sheet from data.

Turn a single avatar/portrait image plus a narration audio file into a lip-synced talking-head MP4, 100% locally with no cloud or API keys. Uses JoyVASA (audio-driven facial-motion diffusion) rendered with LivePortrait, running natively on Apple Silicon (MPS). Works end to end with the other skills: generate the avatar from a description with image-gen and the English/Spanish narration from text with voice-clone-narration, then animate here. Use when the user asks for a talking head, talking avatar, lip-sync video, an avatar that speaks, a spokesperson/presenter clip, or to make a portrait/photo talk or say something from audio or a script.

Turn an mp4 video tour into a 3D Gaussian splat (3DGS) point cloud you can view in the browser, 100% locally on Apple Silicon with no cloud or API keys. Runs the full photogrammetry pipeline - frame extraction (ffmpeg), camera-pose Structure-from-Motion (pycolmap), and Gaussian-splat training (Brush, native Metal via WebGPU) - then compresses the result to web-friendly .sog/.spz with PlayCanvas splat-transform and previews it in a bundled Aholo viewer. Output is compatible with the Aholo viewer (aholojs.dev) and any 3DGS viewer that reads PLY/SPZ/SOG. Use when the user wants to convert a video / room tour / walkthrough / drone flythrough into a gaussian splat or 3DGS scene, mentions aholo, splats, radiance fields, photogrammetry from video, or a "3D capture" from footage.

Clone a voice from a short reference sample - or design a brand-new voice from a text description - and generate expressive MP3 narrations 100% locally (no cloud, no API keys), in English or Spanish, with inflection controls. Uses Chatterbox TTS (MIT) for zero-shot cloning via mlx-audio on Apple Silicon (with a PyTorch fallback elsewhere) and Qwen3-TTS VoiceDesign (Apache-2.0) to invent a voice from a description. Long scripts are sentence-chunked with a consistent voice and encoded to a single MP3. Use when the user asks to clone a voice, make a voiceover/narration/VO for a reel or demo video, do local text-to-speech, turn a script into an mp3, generate a synthetic voice from a description, or say something "in my voice".