Skip to content

Releases

v2.1.0 保留因果链与字幕覆盖验证

Choose a tag to compare

@asdhabdua asdhabdua released this 07 Jul 09:33

What's Changed
新增 scripts/extract_key_sentences.py:从字幕 TXT 中提取含"为什么/原因是/假设/如果...那么/阈值/代价/开销/重传"的因果解释句
升级 scripts/verify_docx.py:新增 --subtitle 参数,自动生成字幕因果句覆盖率检查
升级 scripts/score_frames_concurrent.py 的 --mode extract prompt:保留图中完整因果链,新增 reasoning 字段
同步更新 SKILL.md / CLAUDE.md / AGENTS.md / README.md 的 9 步工作流

B站视频笔记生成Skill--v2.0.0

Choose a tag to compare

@asdhabdua asdhabdua released this 06 Jul 09:39

v2.0.0 发布说明

概述

本版本是项目的重大重构,将其打造为一个通用的、多 Agent 兼容的 B 站视频笔记工具。

核心流程改进

  • 将视觉分析拆分为两个高效步骤:

    1. score 模式 — 对 20-40 张去重后的帧并发打分,选出最终 7-12 张帧
    2. extract 模式 — 只对最终帧并发提取图中文字、公式、表格、概念
    • 这样可以最小化总 token 和 API 成本,并避免任何串行 vision_analyze 调用
  • 为 score_frames_concurrent.py 新增 --mode extract:

    • 同一个脚本,两种用途
    • 通过 ThreadPoolExecutor 并发发起 HTTP 请求

多 Agent 支持

新增专门面向不同 AI Agent 的指令文件:

  • SKILL.md — 供 Hermes Agent 使用
  • CLAUDE.md — 供 Claude Code 使用
  • AGENTS.md — 供 Codex CLI 使用

所有 Agent 文件都基于同一套命令行流程,但按各平台的习惯进行格式调整。

更好的配置方式

  • 将写死的 Sophnet API 配置替换为通用的 OpenAI 兼容变量:
    • VISION_API_KEY
    • VISION_BASE_URL
    • VISION_MODEL
  • 默认示例改为 OpenAI(gpt-4o),同时支持 Sophnet、本地 vLLM 等任意兼容端点

更完善的项目结构

  • 新增 README.md:完整使用指南、徽章、常见问题、目录结构
  • 新增 CONTRIBUTING.md:贡献指南和隐私注意事项
  • 新增 .gitignore:保护 .env、Cookie、视频、帧图、笔记等

隐私与可移植性

  • 从仓库中彻底移除了个人路径、BV 号、API key、本地文件路径
  • templates/env.example 和 bilibili_cookies.txt.example 仅使用占位符

快速开始

git clone https://github.com/asdhabdua/bilibili-video-notes-skill.git
cd bilibili-video-notes-skill
pip install -r scripts/requirements.txt

cp templates/env.example .env
cp bilibili_cookies.txt.example bilibili_cookies.txt
# 填写 API key 和 SESSDATA

python scripts/extract_frames.py BV1xx411c7mD --page 1 --mode cover --subtitle --workspace ./workspace --frames ./frames/p01

完整 8 步流程请参阅 README.md。

v2.0.0 Release Notes

What's New

This release is a major refactor of the project into a universal, multi-agent compatible Bilibili video notes tool.

Core Workflow Improvements

  • Split vision analysis into two efficient steps:

    1. score mode — score 20-40 deduplicated frames concurrently to select the best 7-12 final frames
    2. extract mode — run concurrent vision extraction only on the final frames to get text, formulas, tables, and concepts for note writing
    • This minimizes total token/API cost and avoids any serial vision_analyze calls
  • Added --mode extract to score_frames_concurrent.py:

    • Same script, two purposes
    • Concurrent HTTP requests via ThreadPoolExecutor for both steps

Multi-Agent Support

Added dedicated instruction files for different AI agents:

  • SKILL.md — for Hermes Agent
  • CLAUDE.md — for Claude Code
  • AGENTS.md — for Codex CLI

All agent files use the same command-line workflow, but are formatted for their respective platforms.

Better Configuration

  • Replaced hardcoded Sophnet API settings with universal OpenAI-compatible variables:
    • VISION_API_KEY
    • VISION_BASE_URL
    • VISION_MODEL
  • Default example now points to OpenAI (gpt-4o), but supports Sophnet, local vLLM, or any compatible endpoint

Better Project Structure

  • Added README.md with full usage guide, badges, FAQ, and directory overview
  • Added CONTRIBUTING.md with contribution guidelines and privacy rules
  • Added .gitignore to protect .env, cookies, videos, frames, and notes

Privacy & Portability

  • Removed all personal paths, BV numbers, API keys, and local file paths from the repository
  • templates/env.example and bilibili_cookies.txt.example use placeholders only

Quick Start

git clone https://github.com/asdhabdua/bilibili-video-notes-skill.git
cd bilibili-video-notes-skill
pip install -r scripts/requirements.txt

cp templates/env.example .env
cp bilibili_cookies.txt.example bilibili_cookies.txt
# fill in your API key and SESSDATA

python scripts/extract_frames.py BV1xx411c7mD --page 1 --mode cover --subtitle --workspace ./workspace --frames ./frames/p01

See README.md for the full 8-step workflow.

B站视频笔记生成Skill--v1.0.1

Choose a tag to compare

@asdhabdua asdhabdua released this 30 Jun 12:24

更新内容
质量标准:核心是总结好视频内容,不做数量要求
每个知识点应包含截图+表格+公式+做题要点+WHY解释+助记口诀
提示词增加第0步:git clone + install.py 一键安装
邮箱更新为 EugenegengU@outlook.com
CONTRIBUTING.md 恢复为独立文件

B站视频笔记生成Skill--v1.0.0

Choose a tag to compare

@asdhabdua asdhabdua released this 30 Jun 07:57

功能
🎬 B站视频下载(720p)+ AI字幕获取
📸 每10秒全覆盖抽帧,不丢任何画面
🔍 OCR + 感知哈希双重去重(129帧→20-30帧)
🤖 两轮 Vision 打分:知识点分组 + 选最完整版
📝 融合字幕+截图生成DOCX笔记
🔧 支持 Hermes Agent / Claude Code / Codex / 任意AI