4nrry/media-agent
Media production by an agent on a machine with an 8 GB GPU and commercial licensing, one skill per modality with rules that hold for any model: image and video generation, image retouching, voice-over at -16 LUFS, music and sound with per-item licenses, video composed as code, editing finished video, a 3D product turntable, scripted 3D scenes with characters in Blender, and sourcing rules for factual videos. Two hooks warn: where vipsthumbnail -o really writes, and when Kinocut runs before any skill is loaded.
Builds animated 3D shots as code in headless Blender for a video or a series — scripted sets, low-poly rigged figures from a .glb, a camera that never stops, preview and final quality, cached plates, a frozen-frame check. Use for a 3D scene or animated series in Blender, a character that slides, faces the wrong way or stands in the wrong pose, EEVEE renders that take too long, re-rendering only the shots that changed, frames that came out identical, or 3D too heavy for the browser. A single product spin is in 3d-turntable.
Renders a 360-degree turntable of a .glb or .gltf product model as transparent PNG frames for video, in headless Blender with Cycles on the GPU. Use for a pre-rendered product spin, a device turntable, headless Blender, Cycles, CUDA or OptiX, or Blender segfaulting on devices. Scripted scenes with characters are in 3d-scenes; real-time 3D in the composition is in video-composition.
Gets music, sound effects, stock media and audio fixes for video under commercial licenses — exact-length music, SFX, stock audio, footage and photos with credits, noise removal, captions from real speech. Use for a soundtrack, royalty-free music, SFX, stock footage or photos, CC0 or CC-BY credits, subtitles for a real recording, or a named source (ElevenLabs, ACE-Step, Freesound, Pexels, Pixabay).
Generates images for a marketing piece with any image model — explore, choose on a contact sheet, final still, prompted edits, and what never to generate. Use for a generated image, scene, still, cover, moodboard, storyboard frame or character sheet, or when an image model is named (Grok Imagine, Gemini, Nano Banana).
Retouches images on your own GPU with commercially licensed models — background removal, upscaling, removing a logo or object, masked inpainting, stills generated locally. Use for a transparent PNG, upscale, removing an overlay from your own image, inpainting with a mask, ComfyUI or VRAM limits, or a model like FLUX, Z-Image, RMBG, BiRefNet or LaMa.
Keeps a factual video honest when AI writes, voices and illustrates it — every spoken or shown fact tied to opened sources, rumors and single-source claims kept out, outlets that disagree shown as disagreeing, both sides with equal time, sources on screen, and a check that blocks the render. Use for a news, explainer or true-story video, deciding whether a claim can go in, reviewing a script or episode spec for unsourced claims, numbers that differ between outlets, showing sources on screen, political or election content, or a channel of AI-made videos about real events.
Builds a marketing video as code from its source project — frame-driven scenes, captions from the voice JSON, b-roll placement, ducked music, real-time 3D, and 16:9, 1:1, 9:16, 4:5 plus a carousel from one source. Use for an institutional video, a short vertical video (Reels, Shorts, TikTok), code-drawn motion graphics, LinkedIn formats, declaring AI at upload, Remotion, Three.js, or HTML rendered to video.
Edits a finished video file when only the rendered mp4 exists, with typed operations that return structured results; text or scene changes go back to the source. Use to trim, cut a teaser, merge, resize or crop an existing file, burn in an .srt or subtitles, normalize a finished mp4's audio or extract a frame, or when Kinocut or ffmpeg is named.
Generates short AI video shots with any video model — animate a chosen still, one camera move, resolution by use, check every clip. Use for a generated clip or generated b-roll, an animated photo, image-to-video or text-to-video, or when a video model or route is named (Veo, Omni, Google Flow, Grok Imagine, Sora, Kling). Stock footage is in the audio skill.
Produces edit-ready voice-over with any TTS — one sentence per file, spoken text apart from captions, -16 LUFS, a durations JSON for captions and ducking. Use for narration, TTS, pt-BR speech, loudnorm, captions timed from the voice JSON, translating a voice-over, a quiet or mispronounced voice, or a named TTS (ElevenLabs, Kokoro).