Collect Commerce Material
Turn fragmented material into publishable knowledge while keeping evidence private.
Required flow
Follow this order:
original evidence -> transcription/OCR -> cleaned text -> summary -> tools and platforms -> topic tags -> Library/Case Library
- Create an intake folder with
python scripts/init_material.py --root <materials-root> --title <title> --kind <text|image|video|mixed>.
- Put or reference the untouched material under
raw/. Never publish this directory. Never commit cookies, session tokens, or authenticated page exports.
- For video or audio, invoke
$transcribe-short-video. Save its output under processed/transcript/ and review names, figures, and ambiguous speech.
- For images or screenshots, inspect all images with the model's image-reading capability. Save faithful OCR in
processed/ocr.md; preserve reading order and mark illegible text. Do not invent missing copy.
- For pasted text, preserve the original in
raw/source.txt, then put only normalization fixes in processed/cleaned.md.
- Extract claims, operational steps, tools, platforms, metrics, and caveats. Distinguish what the material states from analyst inference.
- Route with the rules below and write a Markdown record conforming to
references/schemas.md.
- Update
record.json with status, route, tags, and artifact paths. Write frontend data only after the result is reviewed.
Routing
Choose Library for reusable knowledge: methods, tools, platform concepts, terms, SOPs, Listing guidance, advertising, logistics, FBA, profit calculation, SEO/GEO, or workflows.
Choose Case Library only when the material centers on an identifiable brand, shop, product, website, or campaign. A publishable case must include a real site URL and at least one website screenshot. Also collect verified external traffic links such as Facebook, Instagram, TikTok, YouTube, Reddit, or X/Twitter when present. Do not fabricate missing channels.
If material contains both, create one primary record and link smaller reusable insights to it; avoid duplicating the same prose in both libraries.
Public output rules
- Do not expose raw evidence, OCR dumps, transcripts, author identity, or source attribution on the frontend unless the user asks.
- Include concise, actionable conclusions plus verified official links for named tools and platforms.
- Use descriptive titles; do not name an entry after an unknown creator.
- Mark inference internally as
analysis; do not present it as a sourced fact.
- Preserve useful numbers with context, but avoid repeating promotional promises as guaranteed outcomes.
Read references/schemas.md before writing a Library or Case Library record. Read references/taxonomy.md when selecting tags.
1---2name: collect-commerce-material3description: Ingest cross-border ecommerce material from pasted text, local images or screenshots, downloaded post folders, and local video or audio; preserve private raw evidence, invoke transcription or OCR, extract practical knowledge, tools, platforms, and tags, then route the result to Library or Case Library. Use when organizing Xiaohongshu, Douyin, ecommerce notes, tutorials, brand examples, independent-store examples, marketing teardowns, listing advice, logistics guidance, or other fragmented saved material into the user's 出海地图 knowledge base.4---56# Collect Commerce Material78Turn fragmented material into publishable knowledge while keeping evidence private.910## Required flow1112Follow this order:1314`original evidence -> transcription/OCR -> cleaned text -> summary -> tools and platforms -> topic tags -> Library/Case Library`15161. Create an intake folder with `python scripts/init_material.py --root <materials-root> --title <title> --kind <text|image|video|mixed>`.172. Put or reference the untouched material under `raw/`. Never publish this directory. Never commit cookies, session tokens, or authenticated page exports.183. For video or audio, invoke `$transcribe-short-video`. Save its output under `processed/transcript/` and review names, figures, and ambiguous speech.194. For images or screenshots, inspect all images with the model's image-reading capability. Save faithful OCR in `processed/ocr.md`; preserve reading order and mark illegible text. Do not invent missing copy.205. For pasted text, preserve the original in `raw/source.txt`, then put only normalization fixes in `processed/cleaned.md`.216. Extract claims, operational steps, tools, platforms, metrics, and caveats. Distinguish what the material states from analyst inference.227. Route with the rules below and write a Markdown record conforming to `references/schemas.md`.238. Update `record.json` with status, route, tags, and artifact paths. Write frontend data only after the result is reviewed.2425## Routing2627Choose **Library** for reusable knowledge: methods, tools, platform concepts, terms, SOPs, Listing guidance, advertising, logistics, FBA, profit calculation, SEO/GEO, or workflows.2829Choose **Case Library** only when the material centers on an identifiable brand, shop, product, website, or campaign. A publishable case must include a real site URL and at least one website screenshot. Also collect verified external traffic links such as Facebook, Instagram, TikTok, YouTube, Reddit, or X/Twitter when present. Do not fabricate missing channels.3031If material contains both, create one primary record and link smaller reusable insights to it; avoid duplicating the same prose in both libraries.3233## Public output rules3435- Do not expose raw evidence, OCR dumps, transcripts, author identity, or source attribution on the frontend unless the user asks.36- Include concise, actionable conclusions plus verified official links for named tools and platforms.37- Use descriptive titles; do not name an entry after an unknown creator.38- Mark inference internally as `analysis`; do not present it as a sourced fact.39- Preserve useful numbers with context, but avoid repeating promotional promises as guaranteed outcomes.4041Read `references/schemas.md` before writing a Library or Case Library record. Read `references/taxonomy.md` when selecting tags.42
BabyRacoonBBQ