Communitygithub.com

Akira-TL/akira-video-skills

Plan, create, organize, and maintain reusable video assets that are shared across multiple shots; use when character/scene/prop/product prompts, reference images, official materials, identity & state references, or other reusable content is needed.

What is akira-video-skills?

This skill, named 'Video Materials,' is designed for developers and AI agent users who work with video production involving multiple shots. It focuses on managing and maintaining reusable assets stored under `video/materials/`, ensuring that only content genuinely needed across shots is stored there—shot-specific images, sounds, and generation results remain with their respective shots. The skill provides structured guidance for determining what qualifies as a reusable material (e.g., characters, scenes, props, products, visual references), separating identity from state (e.g., costumes, damage levels), and writing prompts alongside reference images. It supports platforms like Claude Code and Cursor, leveraging markdown-based reference documents (`MATERIAL-PLANNING.md`, `REFERENCE-DESIGN.md`, `REFERENCE-CONTINUITY.md`, etc.) to plan batches, handle rights for user-provided or third-party assets, and enforce minimal reference sets to avoid overloading generations. By requiring clear object IDs, roles, and usage boundaries, it prevents duplication and ensures that stable identity/structure info is reused consistently across shots, while shot-specific content stays isolated.

Works with~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/Akira-TL/akira-video-skills/tree/HEAD/skills/video/video-materials

Ask in your favorite AI

Open a new chat with this agent skill pre-loaded.

Documentation

Video Materials

This skill only owns content under video/materials/ that is reused across shots. Shot-specific images, sounds, and generation results stay with their respective shots—do not copy the same material to multiple directories just for categorization.

Before batch-generating reusable assets, read references/MATERIAL-PLANNING.md and work backward from the actual script/shots to determine which long-term materials are needed. Avoid pre-generating just to hit a quantity.

When dealing with user-provided, third-party web, real-person, brand, music, font, or other rights-restricted materials, first read references/SOURCE-RIGHTS.md. Only retain the sources and usage boundaries truly needed for the current project—do not build a complex rights database.

1. Judging Whether Something Is Worth Becoming a Reusable Material

Only content that is referenced across multiple shots, or whose stable identity/structure must be consistently maintained going forward, enters video/materials/.

Common categories include characters, scenes, props, products, and visual references. Category directories are created only when actually needed; prompts and corresponding images can be placed together so that opening the directory gives an immediate understanding of the object.

If a character or scene does not yet have a concrete visual design, first load video-design—do not write vague descriptions directly as image generation prompts.

When creating four-view references for characters/clothing/props/scenes, or unfamiliar mechanical/structural references, read references/REFERENCE-DESIGN.md. For strict four-view templates, see references/FOUR-VIEW-PROMPTS.md.

When the same character, location, product, or prop needs to be consistently reused across multiple generations, read references/REFERENCE-CONTINUITY.md. Use previously approved long-term assets as the baseline reference, rather than re-randomizing identity from text each round.

2. Separating Identity from State

For the same character, prop, or product, first maintain a stable identity, then name distinct states only when there is a real need to differentiate them—for example, Costume A / B, intact / damaged, dry / wet. Do not use a vague "final" asset to represent all states.

Use stable object IDs combined with human-readable purposes in filenames, e.g.:

CHR01_prompt.md CHR01_identity.png CHR01_costume-a.png LOC01_kitchen.png PROP01_damaged.png

Only encode specific states into filenames when the current project genuinely needs to distinguish them simultaneously.

3. Prompts and Reference Images

When writing long-term image prompts, read references/IMAGE-PROMPTS.md. If four-view references are needed, also apply the corresponding strict templates from references/FOUR-VIEW-PROMPTS.md.

Long-term reusable image prompts and their corresponding assets are saved together in the same directory. Prompts describe stable identity, structure, materials, viewpoint, or world rules—they do not carry shot-specific actions or cinematography.

Four-view references for characters, clothing, props, and scenes are recipes made on demand, not a fixed requirement for every project. Once four-view references are needed for an object, use the corresponding strict template—do not arbitrarily switch to free multi-view. For mechanical devices or setups that models cannot easily interpret, add concise mechanism notes next to the assets, specifying fixed parts, moving parts, motion directions, and how a character uses them.

4. Products and Official Materials

Product or brand objects must distinguish official sources from AI-generated/assisted references. Official images, logos, structural documentation, and verified product facts take priority as the basis for structure and authenticity; AI-generated images cannot be reverse-upgraded into product fact sources.

When an advertising project involves product authenticity, load video-advertising. It owns product facts and advertising boundaries. If the current executor cannot load it, pass the capability gap back to akira-video—do not duplicate product fact methodologies inside video-materials.

Before selecting references for a generation task, read references/REFERENCE-ROLES.md to clarify which material is responsible for: character identity, state, scene, product structure, action, cinematography, style, and sound. If the same responsibility has conflicts, resolve them first—do not let the model guess.

5. Minimal Reference Set

When selecting reference materials for a generation task, use the smallest sufficient set that meets the current constraints. Each reference should have a clear purpose (e.g., character identity, costume, scene, product structure, action, or cinematography). Do not dump the entire material library into a single generation just because it "might be useful."

Completion Criteria

Every item in the reusable material directory has a clear object, purpose, and ownership; cross-shot stable information can be directly reused; shot-specific content has not been mistakenly placed into the shared materials area.

Individual skills in this repo

This repo contains 12 individual skills — each has its own dedicated page.

Akira-TL/akira-video-skills

为 Runway 当前视频模型编译和检查生成提示词;当 video-generation 需要使用 Runway(当前主线 Gen-4.5)的 Text to Video 或 Image to Video,并需要 Runway 专用动作、摄影和时序表达时使用。

Akira-TL/akira-video-skills

为 ByteDance / BytePlus Seedance 当前视频模型编译和检查提示词;当 video-generation 需要使用 Seedance 2.5 的参考素材、时间线、同步声音、视频编辑、延长或多镜头能力时使用。

Akira-TL/akira-video-skills

为 Google Veo 视频生成编译和检查提示词;当 video-generation 需要使用 Veo 的 Text to Video、Image to Video、首尾帧、参考元素或同步声音能力时使用,并先核验具体 Google 使用入口与 model 的当前支持范围。

Akira-TL/akira-video-skills

进入并持续推进一个 AI 视频制作项目;读取或建立 VIDEO.md,根据当前视频脚本、复用素材、镜头、生成结果和整片后期状态,按需路由到 canonical 视频专业 Skill,而不是把所有制作规则塞进一个总提示词。

Akira-TL/akira-video-skills

为品牌短片、产品植入剧情、广告或比赛作品建立产品事实与广告边界;当视频需要准确表达某个具体产品的官方功能、参数、结构、使用方式、品牌素材或产品与剧情关系时使用。

Akira-TL/akira-video-skills

设计并维护 AI 视频中跨镜头复用的角色声音、音色参考、对白声音要求与整片声音关系;当同一角色需要跨镜头保持声音一致,或需要规划语音参考、环境声、音效、音乐与画面的职责边界时使用。

Akira-TL/akira-video-skills

设计 AI 视频中的角色与场景视觉方案;当脚本只说明“需要一个什么样的人/场景”,但还没有形成可稳定生成的外形、服装、材质、色彩、空间、光线和视觉识别系统时使用。

Akira-TL/akira-video-skills

组织整片级剪辑、合成、全片声音和最终成片;当多个已采用镜头需要进入 Premiere Pro、After Effects、DaVinci Resolve 等后期工程,或需要整片字幕、Logo、音乐、混音、比例版本和最终导出时使用。

Akira-TL/akira-video-skills

为 AI 图片或视频生成准备可直接使用的提示词与一次性生成包,并在用户把生成结果带回项目后归档;当需要从正式视频脚本、复用素材和 SHOT.md 编译生成内容、选择参考素材、约定返回文件名或清理临时包时使用。

Akira-TL/akira-video-skills

审核 AI 生成的人物/场景等复用素材或镜头生成结果,判断采用、修改提示词、重生成还是留给后期修复;当生成结果返回项目、需要检查身份/结构/连续性/动作/摄影/产品真实性或选择生成结果时使用。

Akira-TL/akira-video-skills

把小说、章节、故事梗概、品牌 brief 或其他上游内容转换成当前视频自己的脚本层;当需要编写或修订 SCRIPT.md、视频人物表演定义、场景定义、对白或镜头总表时使用。

Akira-TL/akira-video-skills

把 SHOTS.md 中的镜头总览展开成可直接制作的单镜头定义;当需要创建或修改 video/shots/<shot-id>/SHOT.md、镜头连续性、摄影、动作、表演、声音或镜头专属生产约束时使用。

Related Skills