镜探 + Avatar Forge：从拆解爆款到数字人口播，Agent Skill 组合完成整条视频创作流水线

视频创作有两个最耗时的环节：把一条好视频真正看懂，以及把一个想法真正演出来。

LycheeAILab 用两个开源 Agent Skill 分别解决了这两件事：

• 镜探 / CineSleuth：把视频交给 Agent，说一句你想分析什么，它自动提取台词、重建场景、拆解镜头，完成全片理解
• Avatar Forge：一张人物图片 + 一段已授权声音 + 一份口播稿，生成完整的数字人口播视频

单独用，各自解决一个问题。组合起来，就是一条从「看懂爆款」到「数字人演绎」的完整视频创作流水线。

---

镜探（CineSleuth）：让 Agent 真正读懂一条视频

它在解决什么

「拉片」是电影圈的专业术语：把一条视频逐帧、逐镜地仔细研究，提取其中的创作逻辑。

做内容的人都知道它有多价值——真正把一条爆款拉透，等于把它的创作方法论学到手。但传统拉片极耗时：要手动打时间码、记台词、标场景、分析镜头语言，一条 3 分钟的视频可能要花 2 小时。

CineSleuth 把这件事交给 Agent 做。

七种分析能力

能力	输入	交付
台词取证	本地视频	逐句台词、说话人、语气、字幕差异与精确时间码
逐镜拆解	想关注的维度（导演/摄影/剪辑）	景别、角度、运动、转场、人物动作与画面文字
场景重建	任意视频	物理场景、镜头和内容段落的清晰区分
声音分析	原始音轨	人声、音乐、环境声、音效、静音与声画关系
结构分析	分析目标	开场钩子、节奏、信息密度、情绪推进与 CTA
镜头提示词	视频中每个画面	逐镜生成可直接用于视频生成的中文提示词
长视频拉片	最长 5 分钟视频	本地智能切分、断点续跑、跨片段合并与完整报告

一句话触发，不需要记命令

使用 CineSleuth 完整拉片这个视频，输出逐句台词、物理场景、逐镜表和视听分析。

使用 CineSleuth 拆解这条短视频的开场钩子、内容结构、节奏、字幕设计和声音设计。

使用 CineSleuth 从导演和剪辑角度分析这部短片，重点说明每个镜头为什么放在这里。

技术架构

视频 → 本地媒体探测 → 语音边界 + 镜头边界 → 智能切片
     → 分段证据 → 全局时间线 → Agent 推理 → 完整拉片报告

关键设计：

• 本地切片：时间码测量和分片在用户设备上完成，不上传原始视频片段
• 断点续跑：已完成的片段保留进度，中断后只处理缺失部分
• 场景不等于切片：技术分段不会自动制造新的场景，跨段场景由 Agent 根据地点、时间、人物连续性合并
• 证据先于结论：每个判断都回到具体时间码，听不清 / 看不清时直接说明，不猜

---

Avatar Forge：让数字人把文案「演出来」

它在解决什么

口播视频需要出镜。这件事有几个障碍：有些人不想出镜、没有专业拍摄条件、需要快速批量产出不同角色的内容。

Avatar Forge 的解法：一张清晰的人物图片 + 一段已授权的参考声音，创建你的专属数字人，然后把任何口播稿变成可以交付的视频。

也可以直接用内置的「丰富公模」快速起步——不需要提供自己的图片和声音。

五种核心能力

能力	输入	交付
丰富公模	文案或成品音频	快速数字人口播视频
专属数字人	一张清晰人物图片	可用于口播的个人数字人
声音克隆	一段已授权参考声音	专属音色与自然口播音频
视频文案再创作	上传视频或有权使用的抖音链接	原始转写稿 + 独立改写稿
创作工具组合	数字人视频与创作目标	接入 HyperFrames 或 ChatCut，完成字幕和剪辑

完整创作流

选择公模 ──────────────────────────→ 数字人口播视频
上传图片 → 专属数字人 ─────────────↗
参考声音 → 专属音色 ──────────────↗
上传/引用视频 → 提取原文案 → 改写新文案 ─(可选)→ 上述任意路径
数字人视频 → HyperFrames / ChatCut → 完整短视频成品

每个步骤都可以独立使用：只需要文案改写、只需要声音克隆、只需要生成一条视频——任意组合。

安全设计

• 费用确认：未经用户明确确认，不提交可能产生费用的生成任务
• 声音授权：只克隆本人声音或已明确授权的声音
• 抖音下载：不读取 Edge/Chrome 登录态，不依赖 Windows DPAPI 或浏览器扩展
• 密钥不下发：第三方服务凭据留在服务端，不进入仓库或日志
• 禁止冒充：不得用于冒充他人或制作违法内容

---

两个 Skill 的组合工作流

用一个具体场景说明两者如何联动：

目标：看到一条爆款口播视频，想做一条同赛道的内容，用自己的数字人表达。

完整流程

Step 1 → CineSleuth 拉片爆款视频
         ↓ 输出：台词、场景结构、逐镜表、节奏分析、开场钩子、镜头提示词

Step 2 → Avatar Forge 「视频文案再创作」
         ↓ 输入：同一条视频（或 CineSleuth 提取的转写稿）
         ↓ 输出：原始转写稿 + 改写后的新文案（保持结构，内容原创）

Step 3 → Avatar Forge 生成数字人口播
         ↓ 输入：改写稿 + 专属数字人（或公模）+ 专属音色（或公模声音）
         ↓ 输出：完整数字人口播视频

Step 4 → HyperFrames / ChatCut 后期包装
         ↓ 字幕、配乐、B-roll、节奏剪辑、最终导出

一句话指令版本（在 Codex 或 WorkBuddy 里）：

第一步：使用 CineSleuth 拆解这条视频的完整结构，输出台词、场景、逐镜表和开场钩子分析。
第二步：使用 Avatar Forge 从同一条视频提取原文案，并改写成 60 秒、自然口语风格的新稿。
第三步：确认改写稿后，使用 Avatar Forge 用我的专属数字人生成口播视频。

每个步骤独立确认，不自动串联——你可以在改写稿环节修改，在生成前调整数字人选择。

更多使用场景

场景 A：快速复刻成功模版 拿到任意一条自己有权分析的行业爆款 → CineSleuth 提取其结构模版 → Avatar Forge 用同样结构生成新内容

场景 B：批量制作不同角色内容 同一份改写稿 → Avatar Forge 切换不同公模 → 同一内容的多个数字人版本

场景 C：纯分析用途 只用 CineSleuth：学习某导演的镜头语言、研究某个广告的信息层次、提取某部短片的叙事节奏

场景 D：纯创作用途 只用 Avatar Forge：已有写好的文案和声音，直接生成数字人视频，不需要拉片

---

安装方式

两个 Skill 都支持 Codex（插件）和 WorkBuddy（Skill）。

CineSleuth（镜探）

Codex 自动安装：

阅读 https://raw.githubusercontent.com/LycheeAILab/cine-sleuth/main/INSTALL.md，
帮我安装或升级 CineSleuth 插件并创建一个新任务。

WorkBuddy： 下载 cine-sleuth-workbuddy-1.0.3.zip，在 Skills 页面上传。

Avatar Forge

Codex 自动安装：

阅读 https://raw.githubusercontent.com/LycheeAILab/avatar-forge/main/INSTALL.md，
帮我安装或升级 Avatar Forge。

WorkBuddy： 下载 avatar-forge-workbuddy-2.2.0.zip，在 Skills 页面上传。

两个 Skill 首次调用都会通过 lab.lycheeai.com.cn 完成授权，授权后本地只保存用户自己的可撤销 Lab API Key，底层服务凭据不会下发到客户端。

---

相关链接

• CineSleuth：github.com/LycheeAILab/cine-sleuth
• Avatar Forge：github.com/LycheeAILab/avatar-forge
• LycheeAILab：lab.lycheeai.com.cn

Two bottlenecks define video creation: truly understanding a great video, and actually producing your own.

LycheeAILab’s two open-source Agent Skills each solve one:

• CineSleuth / 镜探: Hand any video to an Agent, say what you want to understand — it automatically extracts transcripts, reconstructs scenes, breaks down shots, and delivers a full film analysis
• Avatar Forge: One photo + one authorized voice sample + one script → complete digital human video

Used alone, each solves a distinct problem. Used together, they form a complete pipeline: analyze viral content → rewrite → produce with a digital human.

---

CineSleuth: Teaching an Agent to Actually Read a Video

The Problem It Solves

“Film analysis” (拉片) is a professional practice: studying a video frame-by-frame, shot-by-shot, to extract its creative logic.

Any content creator knows its value — truly analyzing a viral video means extracting its production methodology. But traditional film analysis is brutal: manually logging timecodes, transcribing dialogue, labeling scenes, and analyzing cinematographic language. A 3-minute video can take 2 hours.

CineSleuth delegates this to an Agent.

Seven Analysis Capabilities

Capability	Input	Output
Transcript forensics	Local video	Sentence-by-sentence dialogue, speaker ID, tone, caption discrepancies, exact timecodes
Shot-by-shot breakdown	Your analysis dimension (director / cinematography / editing)	Shot scale, angle, movement, transition, character action, on-screen text
Scene reconstruction	Any video	Clear separation of physical locations, shots, and content segments
Sound analysis	Original audio	Voice, music, ambience, effects, silence, audio-visual relationship
Structure analysis	Analysis target	Opening hook, pacing, information density, emotional arc, CTA
Shot prompts	Each frame in the video	Per-shot Chinese prompts ready for video generation
Long video analysis	Videos up to 5 minutes	Local intelligent segmentation, resume on interrupt, cross-segment merging, full report

One Sentence to Trigger, No Commands to Memorize

Use CineSleuth to fully analyze this video — output line-by-line transcript, physical scenes, shot table, and audiovisual analysis.

Use CineSleuth to break down this short video's opening hook, content structure, pacing, subtitle design, and sound design.

Use CineSleuth to analyze this short film from a director and editor perspective — focus on why each shot is placed here.

Technical Design

Video → Local media probe → Speech boundaries + Shot boundaries → Smart segmentation
      → Per-segment evidence → Global timeline → Agent reasoning → Full analysis report

Key design decisions:

• Local slicing: Timecode measurement and segmentation happen on-device; raw footage is not shipped off
• Resume on interrupt: Completed segments are saved; interrupted sessions only process missing parts
• Scenes ≠ slices: Technical segments don’t auto-create new scenes; cross-segment scenes are merged by the Agent based on location, time, and character continuity
• Evidence before conclusion: Every judgment references a specific timecode; unclear audio/visual is flagged directly, never guessed

---

Avatar Forge: Making the Digital Human Perform Your Script

The Problem It Solves

Talking-head videos require someone on camera. Barriers: some people don’t want to appear, professional recording setups are expensive, rapid batch production of multiple personas is impossible.

Avatar Forge’s solution: one clear photo + one authorized voice sample → your own digital human that performs any script you give it.

Or skip the custom assets entirely and use the built-in library of ready-made digital humans for instant production.

Five Core Capabilities

Capability	Input	Output
Stock digital humans	Script or produced audio	Quick digital human talking-head video
Custom digital human	One clear photo	A personal digital human ready for video production
Voice cloning	One authorized reference recording	Custom voice profile and natural talking-head audio
Video script re-creation	Local video or authorized Douyin link	Original transcript + independently rewritten script
Tool integration	Digital human video + creative goal	Hand off to HyperFrames or ChatCut for packaging and editing

Full Creation Flow

Choose stock avatar ─────────────────→ Digital human talking-head video
Upload photo → Custom digital human ─↗
Reference voice → Custom voice ──────↗
Upload/link video → Extract script → Rewrite script ─(optional)→ Any of the above
Digital human video → HyperFrames / ChatCut → Complete short video

Every step is independently usable — only need a script rewrite, only need voice cloning, only need one video — any combination works.

---

The Combined Workflow

A concrete scenario: you find a viral talking-head video in your niche. You want to create your own content with the same structure, performed by your own digital human.

Complete Pipeline

Step 1 → CineSleuth analyzes the viral video
         ↓ Output: transcript, scene structure, shot table, pacing analysis, hook breakdown, shot prompts

Step 2 → Avatar Forge "Video Script Re-creation"
         ↓ Input: the same video (or CineSleuth's transcript)
         ↓ Output: original transcript + independently rewritten script (same structure, original content)

Step 3 → Avatar Forge generates digital human video
         ↓ Input: rewritten script + custom digital human (or stock) + custom voice (or stock)
         ↓ Output: complete digital human talking-head video

Step 4 → HyperFrames / ChatCut for finishing
         ↓ Captions, music, B-roll, pacing edits, final export

One-prompt version (in Codex or WorkBuddy):

Step 1: Use CineSleuth to break down the full structure of this video — output transcript, scenes, shot table, and opening hook analysis.
Step 2: Use Avatar Forge to extract the original script from the same video, then rewrite it into a 60-second, natural conversational style version.
Step 3: After I confirm the rewrite, use Avatar Forge to generate a talking-head video with my custom digital human.

Each step is individually confirmed — you can refine the rewrite before generation, and adjust digital human selection before committing.

More Use Cases

Case A: Quickly template successful formats
Take any video you have rights to analyze → CineSleuth extracts its structural template → Avatar Forge produces new content with the same structure

Case B: Batch produce multi-persona content
Same rewritten script → Avatar Forge switches between different stock avatars → Multiple digital human versions of the same content

Case C: Pure analysis
CineSleuth only: study a director’s cinematographic language, research an ad’s information hierarchy, extract a short film’s narrative pacing

Case D: Pure production
Avatar Forge only: already have a script and voice — generate the digital human video directly, no analysis needed

---

Installation

Both Skills support Codex (plugin) and WorkBuddy (skill).

CineSleuth

Codex auto-install:

Read https://raw.githubusercontent.com/LycheeAILab/cine-sleuth/main/INSTALL.md,
and install or upgrade the CineSleuth plugin for me, then create a new task.

WorkBuddy: Download cine-sleuth-workbuddy-1.0.3.zip and upload it on the Skills page.

Avatar Forge

Codex auto-install:

Read https://raw.githubusercontent.com/LycheeAILab/avatar-forge/main/INSTALL.md,
and install or upgrade Avatar Forge for me.

WorkBuddy: Download avatar-forge-workbuddy-2.2.0.zip and upload it on the Skills page.

Both Skills authorize via lab.lycheeai.com.cn on first use. After authorization, only your own revocable Lab API Key is stored locally — service-side credentials are never sent to the client.

---

Links

• CineSleuth: github.com/LycheeAILab/cine-sleuth
• Avatar Forge: github.com/LycheeAILab/avatar-forge
• LycheeAILab: lab.lycheeai.com.cn