Viral Chaser
Researchviral-chaserViral Chaser is an Agent Skill for research published in the TeamWiseFlow/xiaobei repository on GitHub: a SKILL.md instruction file an AI agent can load to do this task.
下载分析抖音/B站/小红书爆款视频,生成追爆报告。仅产出报告,脚本另外使用 video-product 技能根据 追爆报告.md 生成。
How to use this skill
Bring this guide into your coding agent with a prompt tailored to the tool you use.
- Open your project in Codex.
- Copy and paste the prompt below.
- Review files and risks before approving any action.
Inspect the Agent Skill at https://github.com/TeamWiseFlow/xiaobei/blob/HEAD/crews/main/skills/viral-chaser/SKILL.md and all supporting files. Report provenance, permissions, scripts, dependencies, and security risks. Propose an exact installation into .agents/skills/viral-chaser, but wait for my approval before writing files or running scripts.
Copying this prompt does not install or run the skill. Review third-party files before use.
🔑 前置:开通火山语音模型(仅首次)
本技能的语音转写(ASR)使用火山引擎豆包语音 · 录音文件极速版(资源 ID volc.bigasr.auc_turbo)。即便账号已订购火山 Code Plan,语音模型仍需单独开通,否则调用会返回鉴权/权限错误。
判断是否已开通:直接跑 Step 3 分析器,若 ASR 报错含 status=45xxxxx 或权限相关码,说明未开通,按下面流程开通一次即可。
开通流程(未开通时,根据下面提示并引导用户在火山引擎控制台操作一次):
- 登录火山引擎控制台,左侧控制面板进入 「开通管理」
- 选择 「语音模型」 选项卡
- 找到对应模型(豆包录音文件识别极速版 /
volc.bigasr.auc_turbo),点击 操作 → 立即使用 - 在新打开的页面中点击 「试用」 (点击试用后会赠送20小时,可以先用,后续再点击开通付费)
- 开通后,复制页面下方的 Access token Secret Key,提供给小贝,由小贝写入实例环境变量
环境变量(开通后由小贝配置,用户无需手动设置):
| 变量 | 说明 |
|---|---|
VOLC_ASR_APP_KEY | 火山控制台获取的 APP ID / App Key(必需) |
VOLC_ASR_ACCESS_KEY | Access Token(旧控制台双头鉴权用;新控制台可留空,仅用 VOLC_ASR_APP_KEY) |
VOLC_ASR_RESOURCE_ID | 资源 ID,默认 volc.bigasr.auc_turbo,一般无需改 |
写入流程:用户把
VOLC_ASR_APP_KEY/VOLC_ASR_ACCESS_KEY交给小贝后,小贝应 spawn 一个IT engineer作为 subagent 去把这两个变量添加到实例环境变量中——IT engineer 掌握如何在本机环境变量 / 服务配置里安全添加此类密钥的规范。小贝本人不要直接写环境变量文件。
关于接口选型:火山 ASR 分录音文件标准版 2.0(
volc.seedasr.auc,单价最低,但只接受音频公网 URL,需自备 TOS 对象存储)、极速版(本技能采用,支持本地文件 base64 直传、一次返回)、闲时版(24h 内返回,不适合交互流程)、流式(实时上屏用)。viral-chaser 输入是本地 audio.wav,极速版免托管、原生返回时间戳,综合最合适。若后续为降本要切标准版 2.0,需额外引入 TOS 上传环节。
Viral Chaser(追爆分析 — 报告产出)
Use this skill when:
- 用户提供抖音 / B 站 / 小红书视频链接,希望分析并制作同类视频
- 需要分析爆款视频的结构和公式
本技能仅产出追爆报告,不生成脚本,不制作视频。报告产出后,直接进入 video-product 技能,按追爆报告生成脚本并完成后续生产。
Supported platforms: 抖音(Douyin)、B 站(Bilibili)、小红书(XHS — 仅视频笔记)
Not supported: 微信视频号、TikTok
⚙️ 执行方式(强制)
本技能涉及多步骤生产流程,你应该 self-spawn 一个 subagent 来执行,原因:subagent 独立上下文,不会因对话历史积累而降低输出质量。
你只负责跟进subagent的执行,避免它们长时间卡在某个步骤,必要时可以提供提示或调整执行策略。
Workflow
Step 1 — Create workspace
Before anything else, create the working directory for this video under output_videos/:
VIDEO_SLUG="<platform>-<contentId>" # e.g. douyin-7389abc or bilibili-BV1xx
mkdir -p "output_videos/${VIDEO_SLUG}/references"
All downloaded files, analysis results, and generated reports will be saved under this directory. The references/ subdirectory holds the raw assets (video, audio, key frames) downloaded by the analyzer script.
Step 2 — Check login (skip for public Bilibili videos)
Use the login-manager skill to check the session:
platform:douyin|bilibili|xhs- 探活按 login-manager SKILL.md 步骤 0:
camoufox-cli --session <platform> --persistent --json open <首页>(默认 headless)+snapshot看是否跳登录页(XHS 用xhs-browse) - If exit code 2 (session expired), execute the login flow described in the login-manager skill (原则 3:douyin / xhs-browse 有头手动登录;bilibili 有头登录;login-manager 管的 5 平台之一),then retry 探活
- 登录就位后同时导出 cookie + UA(原则 4):
camoufox-cli cookies export ~/.openclaw/logins/<platform>.json+camoufox-cli identity export ~/.openclaw/logins/<platform>.ua.json
Step 3 — Run the analyzer
Set the output directory to the references/ subdirectory via the OUTPUT_DIR environment variable:
OUTPUT_DIR="output_videos/${VIDEO_SLUG}/references" viral-chaser <url> [--no-frames]
<url>: Full or short-link URL of the video(支持短链,如xhslink.com/o/xxx、v.douyin.com/xxx、b23.tv/xxx,脚本内部跟随重定向解析)--no-frames: Skip key frame extraction (faster, audio-only analysis)OUTPUT_DIR: Must point to thereferences/subdirectory under the workspace created in Step 1
⚠️ exec allowlist 注意:上面这行
OUTPUT_DIR=... ./script是标准 shell 写法,但在 openclaw exec allowlist 下,内联 env 前缀会触发 allowlist miss。通过 exec 工具调用时,请把OUTPUT_DIR放到 exec 的env字段里传,而不是写成内联前缀;同理避免mkdir ... ; echo这类分号复合命令(分号会被当成路径的一部分)。脚本本身已正确读取OUTPUT_DIR落盘,问题只在调用规范。
The script outputs a JSON object to stdout. Read it and proceed with analysis.
Output JSON structure:
{
"ok": true,
"platform": "douyin",
"metadata": {
"contentId": "...",
"title": "...",
"desc": "...",
"author": "...",
"durationSeconds": 89,
"coverUrl": "...",
"stats": { "playCount": 0, "likeCount": 0, "commentCount": 0 }
},
"transcript": {
"text": "全文转录...",
"segments": [{ "start": 0.0, "end": 5.2, "text": "开场文案" }],
"estimated": false
},
"frames": ["output_videos/<slug>/references/frames/frame_00_0s.jpg", "..."],
"localPaths": {
"video": "output_videos/<slug>/references/video.mp4",
"audio": "output_videos/<slug>/references/audio.wav",
"tmpDir": "output_videos/<slug>/references"
}
}
transcript.estimated:false表示segments是火山 ASR 返回的真实时间戳(utterance 级,毫秒精度转秒);true仅在接口异常未返回 utterances 时出现,此时按句切分全文并按字数比例在音频时长上估算分段,时间区间为近似值。正常情况下始终为false。
Exit codes:
0= Success1= Error (URL invalid, download failed, etc.) — report to user2= Cookie expired — execute the browser-based re-login workflow (see login-manager skill), then retry once
Step 4 — Read key frames (if available)
For each path in frames, use the Read tool to load the image and analyze it visually.
Read: output_videos/<slug>/references/frames/frame_00_0s.jpg
Read: output_videos/<slug>/references/frames/frame_01_3s.jpg
...
Analysis Framework
After receiving the JSON output and reading the frames, generate a 追爆报告 in Markdown and save it to output_videos/<slug>/raw_article.md.
1. 内容摘要
1–2 sentences: what core value does this video deliver to viewers?
2. 开头钩子分析(前 0–10 秒)
Based on transcript.segments where start < 10:
- 钩子类型: 提问型 / 冲突型 / 反转型 / 数字型 / 悬念型 / 痛点型 / 利益型
- 具体文案: quote the exact opening line(s)
- 效果评估: why this hook works (or doesn't)
3. 内容结构拆解
Based on transcript segments, divide into logical sections:
| 段落 | 时间区间 | 功能 | 核心内容 |
|---|---|---|---|
| 开场 | 0–Xs | 钩子/引入 | ... |
| 主体一 | X–Ys | 价值/信息传递 | ... |
| 主体二 | Y–Zs | 深化/转折 | ... |
| 收尾 | Z–结束 | CTA/情绪收尾 | ... |
4. 爆款元素评估
Rate each element as 强 / 中 / 弱 with a one-line explanation:
| 元素 | 评级 | 说明 |
|---|---|---|
| 前 3 秒吸引力 | ||
| 痛点共鸣度 | ||
| 悬念设置 | ||
| 情绪触发 | ||
| 价值清晰度 | ||
| CTA 效果 | ||
| 视觉冲击(基于关键帧) | ||
| 节奏把控 |
5. 视觉风格分析(基于关键帧图片)
After reading the frame images:
- 色调风格: 暖色系/冷色系/高饱和/低饱和/黑白
- 构图类型: 人脸近景 / 产品展示 / 场景空镜 / 文字卡片 / 混合
- 字幕/文字覆盖: 字体粗细、位置、是否有背景框、动画感
- 整体视觉标签: 3–5 个关键词(如:「真实感」「强对比」「高信息密度」)
If --no-frames was used or frames is empty, note: "(跳过视觉分析,请重新运行不带 --no-frames 参数)"
6. 可借鉴点
3–5 concise, directly actionable techniques. One sentence each.
7. 目标受众
One sentence describing the primary audience persona.
衔接 video-product
追爆报告产出后,直接进入 video-product 技能流程,并应该明确提示后续工作流程:工作目录为 output_videos/<slug>/,直接按raw_article.md制作脚本.
Notes
- Workspace files are stored in
output_videos/<slug>/— all downloaded assets and analysis reports are kept together. Thereferences/subdirectory contains raw assets from the analyzer. - Bilibili DASH format: if
mediaFormatisDASH, the video and audio streams are separate. The downloadedvideo.mp4contains the video stream only; audio is inaudio.wavafter extraction. This is transparent to the analysis workflow. - XHS video notes only: 小红书图文笔记(image-only)不含视频,viral-chaser 会报错并提示。只有视频笔记(type=video)才能下载和分析。XHS 使用
xhs-browsecookie(消费者端域 www.xiaohongshu.com)。 - ASR segments: 语音转写使用火山引擎豆包语音·录音文件极速版(
volc.bigasr.auc_turbo),原生返回 utterance 级真实时间戳(start_time/end_time,毫秒),脚本转成秒后填入transcript.segments,estimated=false。仅在接口异常未返回 utterances 时,才按句切分全文并按字数比例在音频时长上估算分段(estimated=true)作为兜底。开通/鉴权见文首「前置:开通火山语音模型」。 - Exit code 2 — cookie expired: Execute the login flow described in the login-manager skill(原则 3:douyin / xhs-browse 有头手动登录;bilibili 有头登录),导出 cookie + UA 后重试一次。Do not retry more than once.