agitech-lead-hunter（agitech 领头猎人）

Lead Hunter

Autonomous lead generation that finds, researches, and qualifies prospects daily.

First Run (Onboarding)

If skills/lead-hunter/scripts/config.json has "configured": false, run the onboarding interview before anything else. See references/onboarding.md for the full interview flow.

After onboarding, the config is written and the skill switches to hunt mode.

Hunt Mode (Daily Run)

Step 1: Load Config

Read skills/lead-hunter/scripts/config.json for:

• company - who you are and what you sell
• ideal_customer - size, stage, geography, signals
• sources - where to find leads (industry-specific)
• output - where to put leads (asana, notion, csv, markdown)
• outreach - DM template and personalization rules
• filters - what to skip

Step 2: Scrape Sources

For each source in config.sources:

1. Try web_fetch first (fastest, no deps)
2. If blocked (403/Cloudflare): fall back to scripts/scrape.py which uses Crawl4AI with stealth mode
3. If still blocked: use OpenClaw's managed browser via the browser tool
4. Last resort: use web_search with site:<domain> + freshness filter

Extract from each source:

• Company name
• Funding amount and round type
• Location
• What they do (1-2 sentences)
• Investors (if available)
• Article/announcement URL

Step 3: Filter

Apply config.filters and config.ideal_customer to keep only matching leads:

• Round type matches (e.g., pre-seed, seed)
• Amount in range (e.g., $500K-$10M)
• Geography matches
• Industry/vertical matches
• Not in config.filters.skip_industries

Also deduplicate against scripts/seen.json (persisted list of previously found companies).

Step 4: Research Each Lead

For each qualifying company (max 5 per run to stay fast):

1. Website: web_fetch their site - check team page, product, tech stack
2. Team size: web_search for LinkedIn company page - estimate headcount
3. Key person: web_search for founder/CEO LinkedIn - get name, background, LinkedIn URL
4. Opportunity signals: Flag if no CTO, small team, early product, tech stack match

Step 5: Score & Rank

Score each lead 1-10 based on:

• Team size match (smaller = higher for services, bigger = higher for SaaS)
• Funding stage match
• Tech stack alignment
• Opportunity signals (no CTO, hiring, etc.)
• Recency of funding announcement

Step 6: Generate Outreach

For each lead scoring 6+, generate a personalized DM draft using config.outreach.template with:

• Founder's first name
• Specific observation about their product/company
• How you can help (from config.company.value_prop)
• Soft CTA

Step 7: Output

Depending on config.output.type:

asana:

node skills/asana-pat/scripts/asana.mjs create-task \
  --workspace <workspace_id> \
  --parent <parent_task_id> \
  --assignee me \
  --name "Lead: <Company> - <Round> <Amount>" \
  --notes "<full research + DM draft>"

markdown: Append to leads/YYYY-MM-DD.md with full details per lead.

csv: Append row to leads/leads.csv with: date, company, round, amount, location, url, key_person, linkedin, score, dm_draft

notion: (future - document API integration needed)

Step 8: Update State

• Add found companies to scripts/seen.json for dedup
• Log summary to memory/YYYY-MM-DD.md

Step 9: Report

Output structured summary:

## Lead Hunter Report - YYYY-MM-DD
- Sources scraped: X
- Articles found: X
- After filtering: X leads
- Researched: X
- Qualified (score 6+): X

### Top Leads
1. **Company** - Round $Amount | Score: X/10
   Key person: Name (LinkedIn)
   Signal: [why they're a fit]

Scraping Fallback Chain

The skill uses a tiered approach to handle anti-bot protection:

1. web_fetch - default, fastest
2. scripts/scrape.py - Crawl4AI with stealth (handles most Cloudflare)
3. Browser tool - OpenClaw's managed browser (handles everything but slow)
4. web_search site: query - last resort, gets snippets not full pages

The scrape script auto-manages a venv at scripts/.venv/. First run:

python3 skills/lead-hunter/scripts/scrape.py --check

This creates the venv, installs crawl4ai + playwright chromium. Subsequent runs are instant.

Source Discovery

When the user picks an industry during onboarding, the skill suggests relevant lead sources. See references/sources.md for the industry-to-source mapping.

Users can add custom sources at any time by editing config.sources in config.json.

Rules

• Never send DMs automatically - only draft them
• Max 5 fully-researched leads per run (quality > quantity)
• Always deduplicate against seen.json
• Log every run to daily memory
• If a source is consistently blocked, note it in the report so the user can adjust