references/output-template.md

1. Reusable Viral-Content Formula — distill it into directly applicable topic patterns + a structural template + a visual formula.
2. Segment-by-Segment Breakdown of the Visuals + Voiceover — align them on a timeline and label the visuals/voiceover/hook/pacing/transitions/CTA for each segment + how the two work together.
3. Persona and Content Positioning — persona tags, visual symbols, content matrix, differentiation strategy.

This is an analytical skill that produces Markdown reports, not images. To create a cover, use

yyl-video-thumbnail

. For Xiaohongshu image-and-text posts, use

yuyile-social-card-skill

.

1. Prerequisite checklist (run this first every time, but do not repeatedly bother the user):
   • Run

     bash scripts/check-deps.sh

     .
   • If the self-check fails: do not start the teardown; first help the user prepare the environment. Based on the self-check output, explain what is missing, why it is needed, and what the next command is.
     • Docker is not installed or running → explain that Docker is the runtime environment for the local download APIs, and guide the user to install and open Docker Desktop.
     • Douyin_TikTok_Download_API or XHS-Downloader API is unreachable → if Docker is available, run

       bash scripts/bootstrap-local-apis.sh

       , then rerun

       bash scripts/check-deps.sh

       .
     • ffmpeg / whisper is missing → provide the installation commands from

       INSTALL.md

       , install them, and rerun the self-check.
   • If all required dependencies pass: do not ask the user for confirmation; proceed directly to Step 1.
   • TikHub is optional: during initial setup, or when the target platform requires TikHub (YouTube/Kuaishou/overseas platforms/local API failure) and

     TIKHUB_API_KEY

     is not set, give the user two options: “Configure now” or “Skip for now.” After skipping, continue with the platforms covered by the local APIs; do not block single posts from Douyin/Bilibili/TikTok/Xiaohongshu because TikHub is missing.
   Prerequisites:
   • Required: Docker, Douyin_TikTok_Download_API, XHS-Downloader API, ffmpeg, openai-whisper.
   • Optional: TikHub API key, Jina Reader key.
2. Identify: Determine the platform and granularity (single post/account) from the URL. See

   references/fetch-playbook.md

   for the rules.
   • If the granularity is uncertain, initially process it as a single post and correct it after fetching the page.
3. Prepare the evidence package (prioritize the automated workflow):
   • First run

     bash scripts/prepare-assets.sh "<url>"

     .
   • It automatically creates a working directory, saves

     metadata.json

     , downloads videos/images, transcribes

     audio.txt

     , extracts frames, and generates

     sheet_*.jpg

     /

     xhs_sheet_*.jpg

     .
   • After it finishes, first read

     manifest.txt

     in the working directory and echo one sentence explaining “What I found” (platform/available fields/missing items), then begin the teardown.
4. Data-retrieval fallback (continue when the automated workflow fails; do not stop at the first failure). See

   references/fetch-playbook.md

   for details, endpoints, and commands.
   • ① Platform-local APIs (primary):
     • Douyin / TikTok / Bilibili →

       Douyin_TikTok_Download_API

       :

       curl http://localhost:80/api/hybrid/video_data?url=...

     • Single Xiaohongshu post →

       XHS-Downloader

       :

       POST ${XHS_API_BASE:-http://127.0.0.1:5556}/xhs/detail

       .

       download:true

       downloads image-and-text/video files to

       ${XHS_DOWNLOAD_DIR:-$HOME/.xhs-downloader/Volume/Download}

       .
   • ② TikHub (fallback): Use for YouTube, Kuaishou, overseas platforms, or when the local Xiaohongshu API fails. Requires the

     TIKHUB_API_KEY

     environment variable.
   • ③ Jina Reader: Use for WeChat Official Account articles, regular webpages, and image-and-text-only content as a fallback:

     curl https://r.jina.ai/<url>

     .
   • ④ Guided paste: If the content still cannot be obtained → ask the user to send the copy/subtitles/screenshots. Do not return an error.
5. Required for short videos: download + voiceover transcription + visual-frame extraction (all three are mandatory)
   • Douyin / Bilibili / TikTok / YouTube single post → download the video using

     download_addr

     → use

     ffmpeg

     to extract the audio track → transcribe it with

     whisper --model medium --language Chinese

     → use

     ffmpeg

     for scene detection + rhythm sampling + contact-sheet assembly.
   • Xiaohongshu video post → first download the mp4 with XHS-Downloader → reuse the same

     ffmpeg

     +

     whisper

     + contact-sheet workflow.
   • After extracting the frames, use the Read tool to load

     /tmp/bm/sheet_*.jpg

     into your own context—multimodal image reading is the core evidence for the visual teardown.
   • Xiaohongshu image-and-text post → download the images with XHS-Downloader, assemble a contact sheet in image order, and treat the image sequence as the “visual script”; if there is no voiceover, explicitly mark “No voiceover” and do not invent one.
   • WeChat Official Account article → skip downloading/transcription, but treat the body text and images as evidence.
   • See Section 3 of

     references/fetch-playbook.md

     for command templates.
6. Teardown: Apply the common four-layer framework + seven visual dimensions + platform-specific analysis.
   • Script/structure/topic selection/conversion →

     references/breakdown-framework.md

   • Visual layer (visual hook/shot size/cut rhythm/subtitles/visual symbols/color palette/scene blocking) →

     references/visual-framework.md

   • Single post → deeply analyze this post; the visuals and voiceover must be aligned on the timeline (analyze synergy/complementarity/mismatch/foreshadowing/emphasis).
   • Account → first fetch 5–15 samples and identify recurring patterns (topic selection, structure, visual symbols, posting, engagement), then select 1–2 representative posts for a deep teardown.
7. Output + archive: Generate the report according to

   references/output-template.md

   and save a copy at the same time to

   benchmarks/<platform>-<account-name>/<date>-<title-slug>.md

   (create the directory if it does not exist).
   • Archiving is the foundation of the second phase, the “long-term benchmark library.” Save every result.
   • A short-video report must end with

     ## Full Transcript

     , with the complete contents of

     audio.txt

     pasted verbatim; a summary is not sufficient.
   • An image-and-text/web report must include the complete body text; if there is genuinely no voiceover, clearly state “No voiceover.”
8. Wrap-up: Before delivery, verify that the report ends with the full transcript/body text, then provide “The 3 things most worth copying from this post/account” + one actionable next-step recommendation.

• Start automatically as soon as the self-check passes: Do not ask the user whether to begin. You must complete download → transcription → visual analysis → viral-content teardown → archiving. Do not stop after the evidence package is ready.
• Never come back empty-handed: Work through every fallback tier. If the content truly cannot be fetched, guide the user to paste it; never respond with nothing more than “Fetch failed.”
• Short videos require transcription + image reading: Captions + data alone are not enough. Both voiceover and visual contact-sheet evidence are required; without either, the teardown is incomplete.
• The report must end with the full transcript: For short videos, paste the complete contents of

  audio.txt

  under

  ## Full Transcript

  ; for image-and-text/web content, include the complete body text or state “No voiceover.” This is a hard delivery requirement; do not provide only a summary and do not omit anything.
• Make it reusable, not a restatement: Do not merely repeat what the target said or filmed; distill why they did it and how to apply it.
• Evidence-based: Data and original wording must come from the actual fetched content. Mark unavailable metrics as “Not retrieved”; do not fabricate them.
• First principles: Every conclusion should return to “What problem does it solve for viewers / what emotion does it trigger?” Do not pile on jargon.
• Always archive to

  benchmarks/

  ; the filename must include the platform, account, and date.

• INSTALL.md

  — dependency installation guide (for humans) + common issues.

• scripts/check-deps.sh

  — one-click prerequisite self-check, with installation commands for anything missing.

• scripts/bootstrap-local-apis.sh

  — one-click startup for the local Douyin / XHS download APIs, with installation guidance when Docker is missing.

• scripts/prepare-assets.sh

  — automated evidence-package workflow: download → transcribe → extract frames/assemble contact sheets → generate the manifest.

• references/fetch-playbook.md

  — ★ Read before starting: platform/granularity identification table, four-tier data-retrieval fallback, download + transcription + frame-extraction commands, and wording for paste guidance.

• references/breakdown-framework.md

  — common four-layer teardown + platform-specific adaptations (script/structure layers) + method for extracting viral-content formulas.

• references/visual-framework.md

  — ★ Required for short videos: seven visual dimensions + visual/voiceover timeline alignment + method for extracting visual formulas.

• references/output-template.md

  — three-part deliverable template (already includes a visuals + voiceover alignment table) + archive naming convention.

• Long-term benchmark account library: Based on multiple teardowns accumulated in

  benchmarks/

  , summarize patterns across accounts and time (whose topic categories consistently go viral, industry-wide commonalities, and trackable trends).
• Long-video transcription optimization: Automatically split videos longer than 15 minutes and transcribe them in parallel; or integrate commercial APIs (Feishu Minutes/Tongyi Tingwu).
• Xiaohongshu account-level automation: Integrate self-hosted MediaCrawler for account-homepage sampling, comments, and search; single-post downloads are already covered by XHS-Downloader.