视频分析 爆款视频拆解成提示词 By 灵途 AI

概述

视频理解与复刻提示词生成

Installation and upgrades

Follow https://ailingtu.com/install/skills.md. Do not install or upgrade this Skill from GitHub.

Overview

Use this skill to turn a video into a natural-language replication prompt.

Supported inputs:

• A public TikTok, Douyin (抖音), Xiaohongshu (小红书), WeChat Channels (视频号), YouTube, or Instagram URL.
• A local video file (the script uploads it through /v1/file/upload first, then replicates by file id).
• A material/file already uploaded to Lingtu, referenced by businessId + businessType.

Supported API flows:

• File upload: POST /v1/file/upload, multipart/form-data with form field file.
• Replication prompt: POST /v1/material/analysisTask/stream, server-sent events.

Read references/api.md before changing endpoint paths, request fields, response fields, or streaming parsing. Use scripts/lingtu_video_understand.py for deterministic calls.

Configuration

All authenticated requests use only LINGTU_API_KEY and send it as x-api-key. Installation does not require authentication. If LINGTU_API_KEY is missing when running a task, ask the user to create or manage a key at https://ailingtu.com/api-keys and configure LINGTU_API_KEY in the environment running the Skill before retrying. Never ask for, display, or store their API key in chat. Do not generate account-binding URLs.

缺少密钥时，读取 shared/references/api-key-setup.md，向用户展示其操作系统对应的完整步骤（如何打开终端、可复制命令、粘贴密钥、重启应用及生效范围），不要只说“配置环境变量”。已知系统时只展示对应部分；未知时同时提供 Mac 和 Windows。配置由用户在自己电脑上完成，不代为读取或填入真实密钥。

Use https://api.ailingtu.com as the default base URL unless a future API reference specifies another host.

The independently installed package includes shared/scripts, so authentication, HTTP and upload behavior use the same runtime as the other Lingtu Skills.

Do Not

• Do not call WebFetch / WebSearch / curl against the user-provided video URL (TikTok / Douyin / Xiaohongshu / WeChat Channels / YouTube / Instagram / etc.) to pull web snapshots, search-engine cached pages, OG meta, or third-party metadata sites. The replication endpoint already ingests the source video; adding a web lookup just adds latency and is the wrong data path.
• Exception: only fetch external pages when the user explicitly asks ("先去搜一下这个视频的背景再分析" / "go look up the background first").

Interpreting the output

The replication prompt is a model-generated description, not verified footage or a verbatim transcript. Answer the user's question using it, distinguish the returned description from your own inference, and do not invent missing information. Generated dialogue and timestamps must not be presented as confirmed original speech or frame-accurate timing. If details are absent or uncertain, say so. Stream validation checks transport and content shape; it does not verify factual accuracy or completeness of the generated description.

• If the user asks "这个视频讲了什么 / 在讲什么 / 内容是什么 / 在干嘛"，用 2–4 句中文概括已有信息（主体、场景、钩子、结尾）；不强行补齐不存在的卖点或 CTA。默认不附完整英文提示词，用户需要时再提供。
• 用户要求详细拆解时，按“内容概述、钩子、分镜与节奏、卖点与 CTA、可借鉴做法、不确定信息”组织已有信息；没有依据的部分省略或注明未知。“为什么爆”只能提出假设，结合实际指标和评论再判断。
• If the user asks for tags / 关键词 / 选题归类，从提示词里提取并整理。
• If the user asks "怎么二创 / 怎么改成我的版本"，基于提示词给改写建议；需要生成改编视频时使用 lingtu-content-create。
• If the user only说"给我复刻提示词 / 给我提示词"，那就原样返回提示词，不需要总结。
• 默认语言跟随用户：用户用中文提问就用中文解读，提示词原文（通常英文）保留不翻译。

Workflow

1. Decide the input shape.
   • If the user provides a TikTok, Douyin, Xiaohongshu, WeChat Channels, YouTube, or Instagram URL, send { type: "REPLICATION", url }. Accept full share links and short links (e.g. v.douyin.com, www.douyin.com, xhslink.com, www.xiaohongshu.com, channels.weixin.qq.com).
   • If the user provides a local video file path, upload it first through POST /v1/file/upload (form field file), then send { type: "REPLICATION", businessId: <data.id>, businessType: "FILE" }.
   • If the user references an already-uploaded material/file, send { type: "REPLICATION", businessId, businessType } where businessType is MATERIAL or FILE.
2. Always send type: "REPLICATION" for now. The ANALYSIS mode is reserved and not used by this skill.
3. Use the script to validate SSE data and concatenate streaming results. A non-zero exit means the result is unusable; do not interpret raw diagnostic fragments as a completed analysis.
4. Echo the source (URL, local path, or business id) so the data origin is clear, then respond to the user's actual question per the Interpreting the output rules — summarize when they asked "讲了什么", extract tags when they asked for tagging, return the prompt as-is when they asked for "复刻提示词 / 给我提示词".
5. Downstream usage:
   • For 二创/参考结构生成, use lingtu-content-create with the existing analysis brief; do not repeat completed analysis. For 爆款复刻, use lingtu-video-remake for timestamped transcription, semantic cuts, per-segment redrawing and approval, then merging; a structure brief does not replace timestamped transcription.
   • For 打标/检索, keep the prompt as-is or extract tags from it.
   • For “结合数据分析这个 TikTok/Instagram 视频” / “为什么爆”, call lingtu-social-monitor material for real-time metrics. For “看评论区反馈”, call lingtu-social-comments download on TikTok, Instagram, Douyin, WeChat Channels, or Xiaohongshu. Then combine those results with this skill's replication prompt.

Batch / Concurrency Limit

Do not fan out unbounded parallel calls. Each replication holds an SSE connection until generation finishes (tens of seconds), and uploads consume bandwidth.

• Hard cap: at most 10 concurrent in-flight replicate (or upload) invocations, regardless of input shape (URL, local file, or business id).
• For batch jobs of >10 items, run a worker pool with concurrency ≤ 10 and queue the rest. Default to a smaller pool (3–5) when the user has not asked for "as fast as possible" — it is gentler on the server and on the user's bandwidth.
• Do not retry tighter than once every few seconds on transient failures; if the server returns 429 / 5xx, back off before resuming.
• When the user pastes a long list of videos, confirm the batch size with them before kicking off, and surface progress (e.g. "12 / 100 done") rather than running silently.

Use the built-in batch command (default concurrency: 3):

python3 scripts/lingtu_video_understand.py batch --urls urls.txt --output-dir results --workers 3
# Retry failed items while preserving successful results:
python3 scripts/lingtu_video_understand.py batch --urls urls.txt --output-dir results --workers 3 --resume

urls.txt contains one URL per line. Duplicate URLs are deduplicated. Each URL gets its own JSON file (source, status, task id and prompt on success, or error on failure). Progress goes to stderr; stdout contains the batch summary. Any failed item makes the command exit non-zero. Workers must be 1–10. No automatic request retry is performed; after a transient failure, wait a few seconds before explicitly resuming.

For a single input, use replicate --url "..." --json for structured output. --json and --raw are mutually exclusive. --timeout 300 sets the socket timeout, not a total task deadline. Normal text/JSON is emitted only after stream validation; --raw is diagnostic and may contain incomplete or failed output.

Script Usage

Parse a supported public video URL (TikTok / Douyin / Xiaohongshu / WeChat Channels / YouTube / Instagram):

python3 scripts/lingtu_video_understand.py replicate --url "https://www.tiktok.com/@user/video/1234567890"
# Douyin example:
python3 scripts/lingtu_video_understand.py replicate --url "https://www.douyin.com/video/7123456789012345678"
# Xiaohongshu example:
python3 scripts/lingtu_video_understand.py replicate --url "https://www.xiaohongshu.com/explore/64abcdef0123456789abcdef"
# WeChat Channels example:
python3 scripts/lingtu_video_understand.py replicate --url "https://channels.weixin.qq.com/web/pages/feed?eid=..."

Parse a local video file (auto-upload + replicate):

python3 scripts/lingtu_video_understand.py replicate --file ./clip.mp4

Parse an uploaded material/file:

python3 scripts/lingtu_video_understand.py replicate --business-id 123456 --business-type MATERIAL

Upload only (returns the file id and CDN url, no replication):

python3 scripts/lingtu_video_understand.py upload ./clip.mp4

Print the raw stream (for debugging):

python3 scripts/lingtu_video_understand.py replicate --url "..." --raw