crustdata-supabase（crustdata Supabase 搜索）

CrustData + Supabase Prospect Search

Search for ICP-matching leads using CrustData's People Search API with automatic deduplication against a Supabase database. Each run fetches only net-new leads — existing database entries are excluded at the API level via CrustData's exclude_profiles feature.

Prerequisites (One-Time Setup)

1. Supabase Project

You need a Supabase project with URL and service role key. Add to .env:

SUPABASE_URL=https://xxx.supabase.co
SUPABASE_SERVICE_ROLE_KEY=eyJ...

2. Database Tables

The full database schema lives in tools/supabase/schema.sql. Run it via the Supabase SQL Editor or using the setup script:

python3 tools/supabase/setup_database.py

This creates 6 tables: runs, companies, people, signals, enrichment_log, outreach_log. This skill writes to the people table (dedup by linkedin_url UNIQUE constraint).

3. Verify CrustData Token

Ensure CRUSTDATA_API_TOKEN is set in .env.

Phase 0: Intake

Ask the user these questions to build the CrustData filter config:

ICP Criteria

1. What job titles are you targeting? (e.g., "VP Sales", "Head of Growth")
2. Any titles to exclude? (e.g., "Software Engineer", "Intern")
3. Company size (headcount range)? Options: Self-employed, 1-10, 11-50, 51-200, 201-500, 501-1,000, 1,001-5,000, 5,001-10,000, 10,001+
4. Industry focus? (e.g., Software Development, Financial Services)
5. Geographic region? (e.g., "San Francisco Bay Area", "United States")
6. Seniority level? Options: CXO, Vice President, Director, Experienced Manager, Entry Level Manager, Strategic, Senior
7. Any specific companies to target or exclude?
8. Any other filters? (function, years of experience, recently changed jobs, posted on LinkedIn)
9. How many results do you need? (test: 25, standard: up to 500, full: up to 2,500)

Map Answers to Config

Build the filters array using CrustData's exact filter format. Each filter is:

{
  "filter_type": "FILTER_NAME",
  "type": "in" or "not in",
  "value": ["value1", "value2"]
}

Available filter types: CURRENT_TITLE, PAST_TITLE, CURRENT_COMPANY, PAST_COMPANY, COMPANY_HEADQUARTERS, COMPANY_HEADCOUNT, REGION, INDUSTRY, SENIORITY_LEVEL, FUNCTION, YEARS_AT_CURRENT_COMPANY, YEARS_IN_CURRENT_POSITION, YEARS_OF_EXPERIENCE, COMPANY_TYPE, PROFILE_LANGUAGE, SCHOOL, KEYWORD, FIRST_NAME, LAST_NAME.

Boolean filters (no type/value): RECENTLY_CHANGED_JOBS, POSTED_ON_LINKEDIN, IN_THE_NEWS.

For titles, fuzzy_match: true can be added for broader matching.

Save config:

skills/crustdata-supabase/configs/{client-name}.json

Config JSON structure:

{
  "client_name": "happy-robot",
  "search_config_name": "growth-leaders-sf",
  "icp_segment": "growth-leaders",
  "filters": [
    {
      "filter_type": "CURRENT_TITLE",
      "type": "in",
      "value": ["Head of Growth", "VP Growth", "Director of Growth"],
      "fuzzy_match": true
    },
    {
      "filter_type": "REGION",
      "type": "in",
      "value": ["San Francisco Bay Area"]
    },
    {
      "filter_type": "COMPANY_HEADCOUNT",
      "type": "in",
      "value": ["51-200", "201-500", "501-1,000"]
    },
    {
      "filter_type": "INDUSTRY",
      "type": "in",
      "value": ["Software Development", "Technology, Information and Internet"]
    }
  ],
  "post_processing": {
    "strict_title_and_company_match": false,
    "exclude_names": []
  },
  "max_pages": 20,
  "mode": "standard"
}

Database Write Policy

CRITICAL: Never upsert leads to Supabase without explicit user approval.

This skill writes to the people table. Unwanted writes pollute the database, and cleaning up bad entries across multiple test runs is painful.

Required flow:

1. Run --preview first to show total matching count and cost estimate
2. Run --test to fetch a small sample (~25 leads) — present results to the user before any database write
3. Get explicit user approval before upserting to Supabase
4. Only then run the full pipeline

The agent must NEVER pass --yes on a first run. The --yes flag is only for pre-approved automated runs where the user has already reviewed and confirmed the filters work correctly.

If the user hasn't approved the upsert: Export the CSV and show sample results. Let the user review. Only proceed to upsert after they confirm.

Phase 1: Run the Pipeline

python3 skills/crustdata-supabase/scripts/prospect_search.py \
  --config skills/crustdata-supabase/configs/{client-name}.json \
  [--test] [--yes] [--preview] [--dry-run]

Flags:

• --config (required) — path to client config JSON
• --test — limit to 1 page / 25 results max
• --yes — skip cost confirmation prompt. Only use for pre-approved automated runs, never on first runs.
• --preview — show total matching count only (5 credits)
• --dry-run — show what would happen, no API calls

Pipeline Steps

Step 1: Connect to Supabase — Verify connection, count existing leads.

Step 2: Fetch existing URLs — SELECT all linkedin_url from people table for dedup.

Step 3: Build CrustData payload — Combine user filters with exclude_profiles from Supabase.

Step 4: Search page 1 — Get first 25 results + total_display_count. Show cost estimate.

Step 5: Paginate — Fetch remaining pages with 4s delay between requests (rate limit: 15 req/min).

Step 6: Present results to user — Show sample leads (names, titles, companies) and ask for explicit approval before writing to the database.

Step 7: Upsert to Supabase — Only after user approval. Insert new leads, update existing on conflict (linkedin_url UNIQUE).

Step 8: Export CSV — Write results to output/{client-name}-{timestamp}.csv.

Mode Caps

Cost

CrustData charges 1 credit per profile returned (minimum 5 per request). Preview mode costs 5 credits. Always run --preview or --test first to verify filters before a full run.

Phase 2: Review & Refine

Present results:

• Total matching — how many profiles match the filters in CrustData
• Existing excluded — how many were already in Supabase
• New leads found — net-new profiles fetched
• Email coverage — how many have verified emails
• Top 10 leads — name, title, company preview

Common adjustments:

• Too broad — add more filters (region, seniority, headcount)
• Too narrow — relax title matching (fuzzy_match: true), broaden regions
• Wrong ICP — adjust title include/exclude lists
• Too expensive — lower max_pages or switch to test mode

Phase 3: Output

Supabase (Primary)

New leads are stored in the people table with full CrustData data including email, title, company, location, industry, and headcount. Each person is tagged with client_name, search_config_name, and icp_segment for traceability.

CSV (Secondary)

Exported to skills/crustdata-supabase/output/{client-name}-{timestamp}.csv:

Example Usage

Trigger phrases:

• "Find leads matching [ICP description]"
• "Search CrustData for [titles] at [companies/industries]"
• "Run prospect-search for [client]"
• "Source new leads for [client] — skip duplicates"

Preview (count only):

python3 skills/crustdata-supabase/scripts/prospect_search.py \
  --config skills/crustdata-supabase/configs/happy-robot.json --preview

Test run (review results before approving DB write):

python3 skills/crustdata-supabase/scripts/prospect_search.py \
  --config skills/crustdata-supabase/configs/happy-robot.json --test

Full run (only after user approves test results):

python3 skills/crustdata-supabase/scripts/prospect_search.py \
  --config skills/crustdata-supabase/configs/happy-robot.json

CrustData API Reference

• Endpoint: POST https://api.crustdata.com/screener/person/search/
• Auth: Authorization: Token {CRUSTDATA_API_TOKEN}
• Rate limit: 15 requests/minute
• Max per page: 25 profiles (synchronous)
• Max page number: 100
• Dedup: post_processing.exclude_profiles — array of LinkedIn URLs to skip
• Full docs: Crustdata/Crustdata_documentation/people-docs/realtime-searchapi.md
• Filter reference: Crustdata/Crustdata_documentation/how_to_build_people_filters.md