Imported from Orva-Studio/hance (skills/SKILL.md). Install upstream with npx skills add Orva-Studio/hance --skill skills. Copyright stays with the author.

/hance

A single entry point that routes to one of six subcommands: setup, run, try, refine, batch, ui.

Routing

Pick the subcommand from the user's request. When in doubt, ask.

Runner selection (every subcommand)

Pick one runner per invocation, in order:

1. If command -v bun succeeds → use bunx @orva-studio/hance (preferred — fast cold-start, cached after first fetch).
2. Else if command -v node succeeds → use npx @orva-studio/hance.
3. Else → inform the user that a compiled binary is available for download from GitHub releases, but do not install it for them.

Runner priority: Bun (bunx), then Node (npx), then compiled binary as a last resort. Never install the binary for the user — just let them know it exists if they have neither Bun nor Node. Every subcommand must work on a fresh machine; setup only verifies runtime + ffmpeg and shows examples.

The compiled binary expects a hance-gpu sidecar alongside it; a bare ./hance checkout without it fails with ENOENT … hance-gpu. This is why bunx/npx are preferred — they always work. When developing in this repo, run the source directly: bun run packages/cli/src/cli.ts ….

Hard rules

• Never invent flags. Only use names listed in hance --help or hance preset --help.
• Never run a full video render to explore looks — use hance preview (still frame) for try.
• Windows is unsupported. If uname -s reports Windows, stop and tell the user.

References

• references/preset-index.md — presets/index.json schema and rebuild contract.
• references/compare-page.md — /compare route query params and Edit hand-off.
• references/grading.md — how to judge and dial in a film look: the render→read→adjust loop, what makes a grade read as film, and common artifacts. Read before refine.