Avenox video

Skill avenoxai/avenoxskills/skills/avenox-video

Production agent skills for Claude Code, Cursor, and any SKILL.md harness — Codex fleets, video pipeline, monorepo review bundles, multi-chain explorer.

Install

npx -y skills add avenoxai/avenoxskills --skill avenox-video

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

11 days oldThe repository was created 11 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
12 stars12 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Avenox Studio — local-first YouTube video production pipeline (ROUTER, read first). Use for ANY request to edit, cut, produce, assemble, caption, score, or render a long-form video, or to make motion graphics for one. Wraps open-source tooling (auto-editor, mlx-whisper, MLT/melt, ffmpeg) plus HyperFrames for animated graphics. Triggers: "edit this video", "cut the recording", "make graphics", "extract captions", "render the final", a job name, or anything about the video pipeline. The human is director/quality gate; the agent is the operator.

SKILL.md

5.0 KB, as published. Nobody here has run it

Avenox Studio — operator router

The fast operator guide for a local-first, agent-operated video pipeline. Everything runs on your own machine: no cloud editor, no upload-to-render.

The human directs and approves quality; the agent runs the pipeline.

Setup

export STUDIO_JOBS="$HOME/video/projects"   # heavy media lives here
export STUDIO_ROOT="/path/to/this/repo"     # scripts, templates, brand

Requirements: macOS (hardware encode via h264_videotoolbox; Apple Silicon for mlx-whisper), ffmpeg, python3, melt/MLT, Node (for HyperFrames). Most of this works on Linux with libx264 and a CUDA whisper build substituted in.

Operating principles

Media discipline. Heavy media NEVER in a cloud-synced folder — sync will thrash on multi-GB intermediates and can corrupt in-flight writes. Jobs live in $STUDIO_JOBS/<job>/ (raw/ cut/ graphics/ audio/ outputs/). Your notes system holds only the brain: this system, the brand spec, edit.json plans.
Director loop. Produce a preview (graphics stills + a fast draft render) → send for notes → only then final render. Never ship a final without sign-off. This is the single most important rule; an agent that renders finals unreviewed will burn hours on a rejected cut.
Brand is a hard constraint, not a suggestion. Read brand/frame.md before making any graphic. Define it once and lock it. (The reference implementation is deliberately anti-"AI slop": premium editorial, warm paper
- ink + a single accent, no neon/gradient/glassmorphism/3D-gloss.)
Format: YouTube 16:9 1080p60. Preset in brand/presets/youtube-16x9.json.
Finishing is hybrid. Auto-generate the draft; the same .mlt opens in Kdenlive or Shotcut for hand-finishing. Don't try to automate taste.
Transcription defaults to LOCAL mlx-whisper with whisper-large-v3-turbo — fast, free, and strong on non-English audio. Note that most LLM-routing proxies expose no whisper endpoint; if you go remote, use a dedicated speech API.

Scripts (`scripts/`)

Script	Does
`autocut.sh IN.mp4 [balanced\|aggressive\|conservative]`	silence-cut → `_cut.xml` (Premiere) or `--export` variants
`transcribe.py IN.mp4 PREFIX`	→ `transcript/PREFIX_timed.json` + `_narration.txt`
`mltgen.py edit.json out.mlt --base <job-dir>`	edit-list → MLT project (Kdenlive/Shotcut/melt)
`vrender.sh project.mlt out.mp4 [fast\|quality]`	render (fast = HW draft, quality = CRF18 master)
`grabshot.sh`	clipboard screenshot → disk
`slides2png.sh`	legacy static slides — prefer HyperFrames
`remove-silence.py`	standalone silence pass

The 7 steps

Intake — copy raw → $STUDIO_JOBS/<job>/raw/. Confirm the brief and which segments actually matter.
Rough cut — autocut.sh raw.mov balanced → cut/screen_cut.mp4; transcribe.py for the script. Full recipe → avenox-roughcut skill.
Graphics — HyperFrames. Route via the hyperframes skill → usually motion-graphics (short beats), faceless-explainer (concept stretches), or general-video. Read brand/frame.md first; render animated MP4s into graphics/. Full recipe → avenox-graphics skill.
Assemble — write edit.json (template in templates/edit.json) mixing cut/*.mp4 + graphics/*.mp4 + music → mltgen.py edit.json project.mlt --base <job-dir>.
Captions — transcribe.py → .srt; apply brand/caption-corrections.json (copy it from caption-corrections.example.json — a find/replace map for terms your ASR reliably mangles). Ship as YouTube CC, not burned-in.
Music — bed under everything, sidechain-duck under voice, target ~-14 LUFS. Track attribution in CREDITS.md.
Export — vrender.sh project.mlt draft.mp4 fast → director review → vrender.sh … final.mp4 quality → prune scratch files.

Graphics quality bar

HyperFrames clips must obey brand/frame.md. Prefer type-driven, restrained, weighty motion. If a beat doesn't need motion, a clean static frame is fine — don't animate for the sake of animating.

Reference

Rough cut: avenox-roughcut · Graphics: avenox-graphics
HyperFrames skills: hyperframes (router), hyperframes-cli, hyperframes-animation, hyperframes-creative, motion-graphics, faceless-explainer, general-video
Brand spec: brand/frame.md (fill in from brand/frame.template.md)