avenoxskills

Agent skills built and battle-tested in production by Avenox.

These aren’t demos. Each one runs real work — shipping YouTube videos, driving Codex fleets, packaging codebases for external review — and each carries the gotchas that only show up after something breaks at 2am. That’s the part worth having.

Compatible with Claude Code, Cursor, and any harness that reads SKILL.md-style agent skills.

Install

Copy any skill directory into your agent’s skills folder:

git clone https://github.com/avenoxai/avenoxskills.git
cp -R avenoxskills/skills/codex-fleet ~/.claude/skills/

Copy the whole skill directory, including any scripts/ and references/. The video trio also shares avenox-studio/ (see below).

The skills

Agent operations

External model review

Video production

A local-first, agent-operated YouTube pipeline. Nothing uploads to render.

The three video skills share runtime files in avenox-studio/ — scripts, the edit.json template, and the brand spec:

export STUDIO_ROOT="$PWD/avenox-studio"
export STUDIO_JOBS="$HOME/video/projects"   # heavy media — keep OUT of cloud sync
cp avenox-studio/brand/frame.template.md avenox-studio/brand/frame.md
cp avenox-studio/brand/caption-corrections.example.json avenox-studio/brand/caption-corrections.json

Blockchain

Bring your own assets

Two skills expect files this repo deliberately doesn’t ship:

• avenox-thumbnail needs your own mascot reference in assets/. The mascot is channel identity — yours should be yours. It also needs tool logos, which are third-party trademarks; the fetch recipe is included instead of the files. See skills/avenox-thumbnail/assets/README.md.
• avenox-graphics reads avenox-studio/brand/frame.md, which you create from the template. Lock it early — visual consistency compounds, and changing it mid-channel costs more than getting it slightly wrong at the start.

Requirements

Varies by skill; each SKILL.md states its own.

• codex-fleet — Codex CLI 0.128+, authenticated. Uses bash recipes; Windows uses Git Bash. On Windows/Linux omit the macOS-only caffeinate prefix; the two macOS-only helpers, caffeinate and sips, have bundled PowerShell replacements in skills/codex-fleet/scripts/ that need pwsh 7+
• omp-fleet — omp (@oh-my-pi/pi-coding-agent), authenticated against a provider; budget ~0.5GB RAM per concurrent lane bare, ~1.7GB with a typical MCP set auto-discovered
• gptpro — zip, rsync
• video skills — macOS (hardware encode, mlx-whisper on Apple Silicon), ffmpeg, python3, MLT/melt, Node. Most work on Linux with libx264 and a CUDA whisper build substituted in.
• chainscan — an Etherscan V2 API key; Foundry for cast fallbacks

A note on the gotchas

The sections labelled GOTCHAS are the highest-value part of this repo. A few that cost real hours:

• auto-editor v29 leaks the last --cut-out range as a positional input file. Use ffmpeg’s select filter for content cuts.
• Codex’s greedy -i parse eats your prompt unless you put -- before it.
• An omp lane measures ~1700MB against a Codex lane’s ~108MB — but ~75% of that is MCP servers omp auto-discovers and boots per lane, not the harness (~460MB). Every npx-launched MCP server also keeps a resident npm exec parent, so you pay ~50% extra per server for nothing.
• omp has no exec subcommand — non-interactive is -p. And its model ids fuzzy-match, so an unpinned lane can answer from a metered aggregator instead of your subscription.
• Parallel gpt-image-2 jobs share an image cache and can return duplicate renders — md5 the batch, re-fire dupes solo.
• SVG feTurbulence grain must use a fixed seed or rendered frames flicker.
• auto-editor and npx both need the certifi SSL fix or their downloads fail.

License

MIT — see LICENSE. Use them, fork them, improve them.

Issues and PRs welcome, especially “this broke on my setup” reports.