Supplier Scout

Method: JS DOM Extraction (NOT Screenshots)

Why: Screenshots cost ~2,000 tokens each. JS extraction costs ~200-500 tokens per page and returns structured data. For scouting 10 pages (~400 products), that's 3-5K tokens vs 50-100K.

Data Collection Steps

1. Navigate to search -- tool: browser

   • URL: https://s.1688.com/selloffer/offer_search.htm?keywords=亲子装+[category]&sortType=va_rmdarkgmv30rt
   • Common keywords: 亲子装 (parent-child), 亲子连衣裙 (dresses), 亲子泳衣 (swimwear), 亲子运动套装 (athleisure), 亲子睡衣 (pajamas), 母女装 (mother-daughter)
   • Expected: search results page loads with product cards
   • If captcha blocks: see Error Handling

2. Extract product data (JavaScript) -- tool: browser (JS execution)

   • Execute DOM extraction script on search results page
   • Adapt selectors to current 1688 DOM (changes periodically)
   • Expected: JSON array of {title, price, sales, vendor, link}
   • If selectors return empty: take ONE screenshot to inspect current structure, update selectors, retry

   (() => {
     const items = document.querySelectorAll(
       '[class*="offer-list"] [class*="card"], .sm-offer-item',
     );
     return Array.from(items)
       .map((el) => {
         const title = el.querySelector('[class*="title"], .title')?.textContent?.trim() || "";
         const price = el.querySelector('[class*="price"], .price')?.textContent?.trim() || "";
         const sales =
           el.querySelector('[class*="sale"], [class*="deal"]')?.textContent?.trim() || "";
         const vendor =
           el.querySelector('[class*="company"], [class*="seller"]')?.textContent?.trim() || "";
         const link = el.querySelector('a[href*="detail.1688.com"]')?.href || "";
         return { title, price, sales, vendor, link };
       })
       .filter((x) => x.link);
   })();
   

3. Score vendors -- tool: browser (JS extraction on vendor pages)

   • For each candidate, check vendor page for rating, sales, store age, response rate, stock
   • Scoring: 8-10 Excellent, 6-7 Acceptable, 4-5 Risky (flag), 0-3 REJECT
   • See procedures/vendor-vetting.md for detailed checks
   • If vendor page fails to load: score as 0 with note "unverifiable"

   Factor	0 pts	1 pt	2 pts
   Rating	<4.0	4.0-4.4	4.5+
   Sales	None	Some	Active
   Store Age	<1yr	1-3yr	3+yr
   Response	<80%	80-90%	90%+
   Stock	Low/unclear	Moderate	Plenty

4. Deduplicate -- tool: read

   • Load knowledge/1688-seen-urls.jsonl
   • Filter out any extracted URL already in the file
   • This prevents resurfacing products Francisco already reviewed
   • If file missing: proceed without dedup; create file in step 7

5. Filter candidates -- tool: native

   • Must have: adult AND child sizes in same listing, matching design, vendor score >= 6, estimated margin >= 50%, recent listing
   • Expected: filtered list of viable products
   • If zero candidates pass: widen margin to 40% and re-filter; report relaxed criteria

6. Output ranked table -- tool: write / task card

   • Write ranked table to task card, move card to human lane
   • Expected: table with product name, price, margin estimate, vendor score, link

7. Log seen URLs -- tool: write

   • Append ALL extracted URLs to knowledge/1688-seen-urls.jsonl
   • Outcomes: presented, filtered, approved, rejected
   • Expected: file updated with new entries

Alternative Sourcing

When an existing product's supplier becomes unreliable:

1. Search 1688 for same/similar product (use product image search if available)
2. Compare: price, quality indicators, delivery time
3. Present top 3 alternatives with cost comparison
4. Flag to Francisco for decision

Output Format

Deliverable: Scouting Report

Delivery method: Task card (moved to human lane) + Telegram summary File path: knowledge/scouting/scout-YYYY-MM-DD-[category].md

**Supplier Scout -- [Date] -- [Category]**

**Search:** [keywords used] | **Pages scanned:** [N] | **Products extracted:** [N]
**After dedup:** [N] | **After filtering:** [N]

| # | Product | Price | Est. Margin | Vendor Score | Link |
|---|---------|-------|-------------|-------------|------|
| 1 | [Name]  | [CNY] | ~[X]%       | [X]/10      | [->](url) |

**Rejected (top reasons):**
- [N] below margin threshold
- [N] vendor score too low
- [N] already seen

**Notes:** [any observations about market, pricing trends, etc.]

Success Criteria

• Search executed for target category with correct Chinese keywords
• JS extraction returned product data (>0 results)
• Vendor scoring completed for all candidates above initial filter
• Deduplication applied against knowledge/1688-seen-urls.jsonl
• At least 1 product passed all filters (or relaxed criteria documented)
• Ranked table delivered to human lane
• All extracted URLs logged to seen-urls file
• Telegram summary sent to Francisco

Error Handling

Evidence Standards

• Every product recommendation must include direct 1688 URL
• Prices must note currency (CNY) and extraction timestamp
• Margin estimates must show calculation: (retail price - total cost) / retail price
• Total cost must include: product + estimated shipping + platform fees
• Vendor scores must show individual factor breakdown, not just total
• Use confidence tags: [verified] (data from page), [estimated] (calculated), [unconfirmed] (parsing uncertain)
• Flag any product with incomplete data (missing sizes, unclear pricing tiers)
• Distinguish between "minimum order" price and "single unit" price

Permission Tiers