Video / Audio / Streaming Engineer — SEVENTHWELL, Japan Building tools for real-world event production. AI × FFmpeg × Automation
I build agent skills and production software: tools that let AI coding agents (Claude Code, Cursor, Codex) do real media work — measuring first, editing deterministically, and verifying the result before calling it done. By day I run video/audio/streaming on live event floors; the software below comes from that work.
ffmpeg-skill — ★2,700+ · 200+ forks
A local video editor for coding agents: 42 typed FFmpeg tools with a machine-readable contract and a contract-derived MCP server. The workflow is fixed — probe → edit → check → verify — so the agent never guesses frame rates, codecs or container limits.
- Cut, join, silence removal, aspect fitting, captions & word-timed karaoke, overlays, motion graphics
- HDR10 / HLG / Dolby Vision → SDR, 3D LUTs and primary grading
- Audio cleanup, typed dynamics, EBU R128 two-pass loudness, multicam sync with drift correction
- Delivery checks against YouTube / Shorts / Reels / TikTok / X / LinkedIn specs, with fixes
- Project rendering from a declarative file, folder-wide batch with content-hash caching
- No API keys, no cloud, Python standard library only · Install:
npx ffmpeg-skill - 50 days on GitHub Trending (peak #46) · Reviewed by tech media in China, Italy and Japan (note)
Each skill is a self-contained package with a machine-readable contract, a doctor capability check, and verification built in. They compose: most execute through ffmpeg-skill.
- video-editing-skill — Executes typed video-edit requests (
video-editing/request@1) deterministically via ffmpeg-skill and validates the output against the request. - remotion-ffmpeg-skill — Renders measured Remotion motion graphics onto real recordings via ffmpeg-skill; every step verified before the next.
- motion-graphics-skill — Deterministic motion-graphics rendering onto video: lower thirds, picture-in-picture, chroma key, timeline validation.
- subtitle-skill — Generates SRT/WebVTT subtitles deterministically and burns them into video, with font, outline and placement control.
- transcription-skill — Local speech-to-text with faster-whisper; validated transcripts with a defined hand-off format for editing skills.
- color-grading-skill — Typed colour-grading operations executed via ffmpeg-skill, capability-checked against the local FFmpeg build.
- audio-production-skill — Executes typed audio operation graphs (cleanup, dynamics, mixing, loudness) via ffmpeg-skill.
- media-analysis-skill — Measures media files locally with ffprobe/ffmpeg and returns structured observations other skills can plan on.
- thumbnail-skill — Renders typed thumbnail documents into validated PNG/JPEG images.
- imagemagick-skill — Local, non-destructive image editing for coding agents with ImageMagick, plus an MCP interface. HEIC/WebP aware.
- hyperframes-skill — Renders HTML scenes to verified video via HyperFrames and ffmpeg.
- blender-skill — Headless Blender pipeline for agents: inspect, convert, optimize and render 3D assets locally.
- qc-skill — Deterministic QC of video, audio, subtitle and delivery artifacts against typed rules.
- artifacts-skill — Locally verifies AI-generated documents and produces evidence-backed receipts.
- AI-video-production-OS — Open contracts and shared infrastructure for AI-native video production across independent skills.
- video-production-agent — Orchestrator that plans media work and executes it through ffmpeg-skill and the skill family.
- design-check-skill — Audit-only: checks Apple app screens and specs against HIG checklists (iPhone / iPad / Mac) in an audit loop until gaps close. It audits; it does not generate designs.
- paddock — Conference reception system: monitors NAS speaker folders read-only, manages reception intake, delivers materials to venues. Used at real academic conferences.
- Remain monitors — read-only remaining-playback-time monitors with a browser UI: disguise, QLab, PIXERA, Grass Valley T2/T3, ProVideoPlayer
- conductor — Button-deck show control from iPad/iPhone: video switchers, media servers, Stream Deck (drivers).
- crosspoint — iOS remote for Roland V-160HD / V-80HD video switchers over LAN, USB or Bluetooth MIDI.
- timer — Event countdown timer for Windows: separate display and control windows, Rust core.
- sw-ars — Realtime audience polling and quizzes for live venues (PULSE), with add-ons for Slides / PowerPoint / Zoom / Teams.
- voicebooth — Small vocal DAW for song covers: sing over an off-vocal track, check pitch against the guide, export WAV. Free download (Windows / Mac).
- sw-audio — 132 SEVENTHWELL audio plugins as CLAP, with VST3 and AU builds.
- model_600 — Auto-mixing audio suite for OBS Studio: five units for voice, calls, game audio, music and master.
- bdmvkit — Blu-ray (BDMV) authoring: titles, chapters, menus, subtitles.
- vtskit — DVD-Video authoring: writes VIDEO_TS discs with titles, menus, subtitles.
- udfkit — UDF writer for DVD-Video and Blu-ray disc images.
- dvdcc — Embeds line-21 closed captions (CEA-608) into DVD MPEG-2 video without re-encoding.
- oncemore — Mac app that turns videos into DVD/Blu-ray disc images with menus, chapters and subtitles.
- fresnel — Liquid Glass icon composer for layered iOS / macOS / watchOS app icons.
- zippeek — Mac app to compress, extract, and peek inside archives without extracting them.
- Measure, don't guess — probe the actual file before deciding anything
- Verify before "done" — every workflow ends with a check, not a claim
- Names match what the thing actually does — no over-broad claims
- Web: seventh-well.com
- X: @kebura_P
- LinkedIn: in/seventhwell