Communitygithub.com

Gavin152/promo-clip

Make the characters, backdrops, props and screens for a promo-clip video — AI-generated in the house style or processed from your own visuals.

promo-clip とは?

promo-clip is a Claude Code agent skill that make the characters, backdrops, props and screens for a promo-clip video — AI-generated in the house style or processed from your own visuals.

対応✓Claude Code~Codex CLI~Cursor
npx skills add https://github.com/Gavin152/promo-clip/tree/HEAD/skills/art

お気に入りのAIに質問する

このエージェントスキルを事前に読み込んだ状態で新しいチャットを開きます。

ドキュメント

/promo-clip:art

Read ${CLAUDE_PLUGIN_ROOT}/skills/promo-clip/protocol.md first, and the images section of ${CLAUDE_PLUGIN_ROOT}/skills/promo-craft/providers.md. stage art start.

1. Art list

From every Picture: line in V/script.md and the ## Visuals and ## End card sections of V/brief.md, list each asset once:

keykindsourceused in
  • kind — subject (character, prop, object: a die-cut sticker drawn with art()), backdrop (full-frame scene behind a whole scene, drawn with backdrop()), screen (app screenshot shown in phone()), photo (real photo, e.g. in a polaroid()), logo/badge/qr.
  • source — ai, or the user's file.
  • key — short snake_case; it becomes the manifest key the scenes use.

Small things the collage style draws in code need no art: sun, moon, clouds, gulls, phone frames, polaroid frames, speech bubbles, icons (wifi, check, cross, stopwatch, gear, trophy, padlock, credit card, keypad), stamps, confetti, data tags, stick-figure people (person()), boats, islands, compass, maps. List only what must be pictured. Reuse one character pose across scenes where the script allows; each AI image costs ~$0.07.

Show the list and its estimate (AI count × the price from art generate's log) before generating.

2. Make each asset

  • AI subject: art generate <key> "<what, pose, expression, clothing, colours>" then art process <key>. Add --ref <file> (repeatable) for a style or likeness reference from assets/ref/. One subject per image, full body for characters.
  • AI backdrop: art generate <key> "<place, light, mood>" --kind backdrop then art process <key> --kind backdrop (fills the primary format's aspect; JPEG).
  • Own subject cut-out: on a plain magenta background → art process <file> --out <key>; anything else → ask for a cut-out PNG, or use it as a framed photo instead.
  • Own photo / screenshot: art process <file> --nokey --out <key> (just resized), or copy it into assets/img/ and run art manifest.
  • QR / logo / badges: art qr <url> (done in discovery if the end card has a QR); logo and badge files in assets/img/ + art manifest.

After each art process, look at the result (Read the PNG): a stray ground line under the subject, a cut-off limb or a magenta fringe means regenerate (edits can't fix geometry — change the prompt and generate again).

Done when every row of the art list has a file in manifest.json.

3. Check

  • art check — warns when an asset would look soft at the largest output size; re-process with the --max it suggests, or regenerate with --hires=true for 4K.
  • art contact → .work/contact_sheet.jpg. Read it.

Gate: the contact sheet + the art list. Gate per the protocol; a rejected asset is regenerated (new prompt) or replaced. Then stage art done (scenes go stale on a re-run).

Individual skills in this repo

This repo contains 8 individual skills — each has its own dedicated page.

関連スキル