Video Thought Digest 是一个与模型无关、以证据为先的 Agent Skill，用于把链接视频或本地视频整理成可长期保存、可重复分析的研究资料包。
它不会把“资料提取”和“观点解释”压缩成一份混合摘要，而是分别保存原始证据、结构化主张、不同模型的独立评论，以及跨模型综合。这意味着同一段视频以后可以交给其他模型重新审阅，不必重复下载和转写，也不会让新的审阅者受到已有观点干扰。
这不是一键生成视频摘要的轻量工具。
它适合需要保留原始证据、追溯观点来源，并让多个 AI 相互独立审阅的研究、调查与分析工作流。
如果你只想快速了解一段视频的主要内容，使用普通转写或摘要工具会更简单；如果你关心可复核性、模型意见隔离和长期资料留存，这套额外流程才有价值。
来源与溯源信息
-> 保留的媒体文件与 SHA-256
-> 带时间戳转写稿与清洁转写稿
-> 结构化主张与关键帧
-> 中立审阅包
-> 各模型独立评论
-> 跨模型综合
每条资料使用 Markdown、SRT、JSON，以及普通音视频或图片格式保存。manifest.yaml 使用严格的 JSON 语法写入（JSON 是 YAML 的子集），因此无需专有数据库或额外 YAML 依赖。
- 证据与解释彼此分离。
- 各模型独立审阅，在综合阶段之前互不影响。
- 使用媒体哈希识别重复来源或跨平台转载。
- 局部步骤失败时保留已完成产物，并明确记录缺失内容。
- 可通过宿主 Agent 现有工具处理网页链接或本地文件。
- 可选用 Downie 加速 macOS 上的视频获取，但整个流程不依赖 Downie。
- 资料库位置必须显式指定，或通过
VIDEO_THOUGHT_DIGEST_ROOT配置。
skills/video-thought-digest/
├── SKILL.md
├── agents/openai.yaml
├── references/artifact-schema.md
└── scripts/
├── create_entry.py
├── attach_media.py
├── build_review_pack.py
└── validate_entry.py
克隆仓库，然后把 Skill 文件夹复制到你的 Agent 所使用的 Skill 目录。
git clone https://github.com/observerforest/video-thought-digest.git
mkdir -p "$HOME/.agents/skills"
cp -R video-thought-digest/skills/video-thought-digest "$HOME/.agents/skills/"git clone https://github.com/observerforest/video-thought-digest.git
cp -R video-thought-digest/skills/video-thought-digest ~/.claude/skills/其他支持 SKILL.md 约定的 Agent，也可以把同一文件夹安装到各自对应的 Skill 目录。
OpenAI 建议使用插件来进行面向广泛用户的分发。本仓库目前以独立 Skill 的形式发布，方便直接检查和本地安装；以后可以在不改变证据格式的前提下封装为插件。
创建资料条目时可以显式指定根目录，也可以设置环境变量：
export VIDEO_THOUGHT_DIGEST_ROOT="$HOME/Documents/video-research"Skill 不会默认使用某个私人 Obsidian Vault 或其他个人文件夹。
可以直接对 Agent 说：
使用 video-thought-digest 归档并分析这个视频。保留一份中立的证据包，并且不要参考已有评论，独立写出你的审阅意见。
也可以直接运行确定性的条目初始化脚本：
python3 skills/video-thought-digest/scripts/create_entry.py \
--root "$VIDEO_THOUGHT_DIGEST_ROOT" \
--platform local \
--creator "Example Creator" \
--title "Example Video" \
--source-id "demo-001"examples/ 中提供了虚构的来源与主张示例。完整的归档格式约定见 artifact-schema.md。
- 随附脚本需要 Python 3.9 或更高版本。
- 附加媒体文件时，
ffprobe可提高跨平台时长识别能力。 - 在可用时，macOS 的
afinfo会作为备用方案。 - 宿主 Agent 可以使用适配对应平台的浏览器工具、转写引擎或
ffmpeg获取来源内容。 - Downie 仅为可选集成，运行时必须先检查它当前是否存在并可用。
Video Thought Digest 与 Downie 或 Charlie Monroe Software 没有隶属或合作关系。它不会购买软件、绕过访问控制、代解验证码，或规避平台身份验证。
本仓库只包含工作流代码和虚构示例。请勿提交下载的媒体、私人转写稿、Cookie、会话数据、API 密钥或个人评论资料库。
初始化脚本会移除 URL 中的用户信息和片段，并遮盖常见的凭证类查询参数。分享前应优先使用规范的公开来源 URL，并检查生成的 manifest。
仅获取和保留你有权使用的资料。遵守平台条款及适用的版权规则仍是使用者的责任。
运行仅依赖 Python 标准库的测试套件：
python3 -m unittest discover -s tests -v本项目由 observerforest 发起并定义需求、工作流与验收标准。初始代码、测试和文档主要由 OpenAI Codex 的 GPT-5.6-sol 生成，经 Claude Fable 5 和独立 Codex subagent 审阅。
项目发起人未手写代码；所有生成内容均经过测试、审查和人工决策后发布。
Video Thought Digest is a model-neutral, evidence-first Agent Skill for turning linked or local videos into durable research packages.
Instead of collapsing extraction and interpretation into one summary, it keeps source evidence, structured claims, independent model reviews, and cross-model synthesis in separate artifacts. A video can therefore be reviewed again without repeating acquisition or contaminating a new reviewer with earlier opinions.
This is not a lightweight, one-click video summarizer.
It is designed for research, investigation, and analysis workflows that require preserved evidence, traceable claims, and independent review by multiple AI models.
If you only need a quick overview, a conventional transcription or summarization tool will be simpler. The additional workflow becomes valuable when reproducibility, review isolation, and durable evidence matter.
source provenance
-> retained media and SHA-256
-> timestamped and clean transcripts
-> structured claims and keyframes
-> neutral review pack
-> independent model reviews
-> cross-model synthesis
Each entry uses portable files such as Markdown, SRT, JSON, and ordinary media or image formats. manifest.yaml is written using strict JSON syntax (JSON is a subset of YAML), so no proprietary database or additional YAML dependency is required.
- Evidence and interpretation remain separate.
- Reviews are isolated until synthesis.
- Media hashes identify duplicate or cross-posted sources.
- Partial failures preserve completed artifacts and record what is missing.
- URLs and local files are supported through the host agent's available tools.
- Downie can be used as an optional macOS acquisition accelerator, but the workflow does not require it.
- Archive location is explicit or configured through
VIDEO_THOUGHT_DIGEST_ROOT.
skills/video-thought-digest/
├── SKILL.md
├── agents/openai.yaml
├── references/artifact-schema.md
└── scripts/
├── create_entry.py
├── attach_media.py
├── build_review_pack.py
└── validate_entry.py
Clone the repository, then copy the Skill folder into the directory used by your agent.
git clone https://github.com/observerforest/video-thought-digest.git
mkdir -p "$HOME/.agents/skills"
cp -R video-thought-digest/skills/video-thought-digest "$HOME/.agents/skills/"git clone https://github.com/observerforest/video-thought-digest.git
cp -R video-thought-digest/skills/video-thought-digest ~/.claude/skills/Other agents that support the SKILL.md convention can install the same folder in their corresponding Skill directory.
OpenAI recommends plugins for broadly installable distribution. This repository currently publishes the standalone Skill for direct inspection and local installation; it can be packaged as a plugin later without changing the evidence format.
Choose an archive directory explicitly when creating an entry, or set:
export VIDEO_THOUGHT_DIGEST_ROOT="$HOME/Documents/video-research"The Skill never assumes a personal Obsidian vault or another private folder.
Ask the agent:
Use video-thought-digest to archive and analyze this video. Preserve a neutral evidence package and write your review independently from any existing reviews.
The deterministic entry initializer can also be run directly:
python3 skills/video-thought-digest/scripts/create_entry.py \
--root "$VIDEO_THOUGHT_DIGEST_ROOT" \
--platform local \
--creator "Example Creator" \
--title "Example Video" \
--source-id "demo-001"See examples/ for synthetic source and claim data. See artifact-schema.md for the complete archive contract.
- Python 3.9 or newer is required for the bundled scripts.
ffprobeimproves cross-platform duration detection when attaching media.- macOS
afinfois used as a fallback when available. - The host agent may use platform-aware browser tools, transcription engines, or
ffmpegfor source extraction. - Downie integration is optional and must be checked for current availability at runtime.
Video Thought Digest is not affiliated with Downie or Charlie Monroe Software. It does not purchase software, bypass access controls, solve CAPTCHAs, or circumvent platform authentication.
The repository contains workflow code and synthetic examples only. Do not commit downloaded media, private transcripts, cookies, session data, API keys, or personal review archives.
The initializer strips URL userinfo and fragments and redacts common credential-bearing query parameters. Prefer canonical public source URLs and inspect manifests before sharing them.
Only acquire and retain material you are permitted to use. Platform terms and applicable copyright rules remain the user's responsibility.
Run the standard-library test suite:
python3 -m unittest discover -s tests -vThis project was initiated by observerforest, who defined its requirements, workflow, and acceptance criteria. The initial code, tests, and documentation were primarily generated by OpenAI Codex using GPT-5.6-sol, with additional review by Claude Fable 5 and an independent Codex subagent.
The project initiator did not manually write the code; all generated outputs were tested, reviewed, and released through human-directed decisions.