noique/cross-border-ecommerce-skills/contact-extractor

Contact Extractor

Multi-source email extraction with confidence tiering.

Quick Start

python3 scripts/extract_contacts.py linktree_expanded.csv --out linktree_with_contacts.csv
python3 scripts/extract_contacts.py creators.csv --verify --out enriched.csv  # SMTP MX probe

Sources tried (in order)

1. personal_site root + /about + /contact + /press + /work-with-me

• mailto: links → high confidence
• Visible text emails → medium

1. YouTube Data API v3 (set YOUTUBE_API_KEY env var)

• Channel description email → high

1. Apple Podcasts public API (no key needed)

• RSS feed iTunes:owner email → high

1. Email pattern guess (only if 1-3 produce no hits)

• hello@, hi@, contact@, {firstname}@ on personal_site domain
• Low confidence unless --verify runs SMTP MX probe

Required input columns

• personal_site — domain (e.g. nurtuher.com)

Optional input columns

• youtube — channel slug or ID (for YouTube source)
• podcast_apple — Apple Podcast ID (for podcast RSS)
• page_title / linktree_handle — used as name hint for pattern guessing

Output

Same CSV + 5 new columns:

• contact_email_1, contact_email_2, contact_email_3 — ranked
• contact_sources — e.g. "site / podcast" (which sources hit)
• contact_confidence — high / medium / low / none

Confidence rubric

NEVER_DIG_DOMAINS

The script skips visiting known third-party domains that produce noise emails (link shorteners, affiliate platforms, scheduling tools, etc.). See list in source.

Pilot run (2026-05-06, 45 Linktree-derived KOLs)

After fixing personal_site identification (handle-match scoring) + NEVER_DIG filtering:

• 5 high (real mailto / podcast owner / brand support)
• 4 medium (text emails on creator sites)
• 11 low (pattern guess on validated personal_site)
• 25 none (no findable personal_site — recommend IG bio extraction next)

Limitations

• Requires personal_site in input — for KOLs without one (~50-70% of Linktree pool), this skill produces "none". Use IG bio extraction (Apify or manual) for those.
• Pattern guess at low confidence has ~60% land rate (推算 industry baseline) — recommend SMTP probe or Hunter.io verification before mass send.
• YouTube channel businessEmail is hidden behind CAPTCHA — only emails embedded in description are captured.