Clay sourcing → HubSpot → pipeline (3-variant split)

Clay is MCP-only (no REST key), so the contact-finding runs through the Clay MCP and must be
driven by you (Claude). HubSpot writes + the pipeline are handled by Python scripts in
.claude/skills/sdr-pipeline/scripts/. Run from the project root.

Inputs

• A HubSpot company list ID (its members are companies), OR a list of company domains directly.
• Optional: a name for the new HubSpot list, and how many companies to process this run.

Steps

1. Get company domains. If given a HubSpot company list ID:
   python3 .claude/skills/sdr-pipeline/scripts/hubspot_company_domains.py <list_id>
   → JSON [{company_id, name, domain}]. Drop entries with domain: null (log them). Clay needs
   domains, not names. Cap at ~25 companies per run unless the user asks for more (Clay credits).

2. Find contacts via the Clay MCP — one call per domain. Use
   mcp__48bc72f4-64a3-448c-bf77-4f473c7b118a__find-and-enrich-contacts-at-company with:

   • companyIdentifier: the domain (e.g. "acme.com").
   • contactFilters.job_title_keywords: ["CRO","Chief Revenue Officer","VP of Sales","Head of Sales","Director of Sales","SDR Manager","BDR Manager","Sales Development Manager","Director of Revenue Operations","Revenue Operations","RevOps","Sales Operations","Sales Ops"] (values OR together).
   • contactFilters.locations: ["United States"].
   • contactFilters.job_title_exclude_keywords: ["Intern","Assistant"].
   • dataPoints.contactDataPoints: [{ "type": "Email" }] (we need work emails to enroll).
     The tool is async: it returns a taskId. Poll with
     mcp__48bc72f4-64a3-448c-bf77-4f473c7b118a__get-task-context until the contacts + emails are
     resolved before reading results. Collect each contact's name, title, email, company, domain, and
     LinkedIn URL.

3. Assemble candidates into a JSON file, e.g. /tmp/clay-candidates.json — a flat list of
   {first_name,last_name,title,email,company,domain,linkedin_url} objects (the script also accepts
   firstName/jobTitle/linkedin key variants).

4. Dedup + create + list + ingest (one command, deterministic):
   python3 .claude/skills/sdr-pipeline/scripts/source_contacts.py /tmp/clay-candidates.json --list-name "AI SDR Sourced <date>"
   This: drops no-email + duplicate rows, keeps only ICP titles (via buyer_group, assigning persona),
   dedups against HubSpot by email (drops anyone already there), assigns each net-new contact a
   variant round-robin (value-give / earn / show), creates the net-new contacts in HubSpot,
   makes a static HubSpot list and adds them, writes data/outreach/sourced-<list>.jsonl, and runs
   sdr_batches init so they land in the pipeline as pending batches with their variant set.
   (Add --no-ingest to stop before the pipeline, or --no-hubspot to dry-run the jsonl only.)

5. Report the summary it prints: candidates, ICP, already-in-HubSpot, net-new, per-variant counts,
   the new HubSpot list id, and the new pending batches. Tell the user to open the Pipeline tab and
   Generate copy (the per-contact variant is honored automatically — each contact is written with
   its assigned instruction set) then enroll (each routes to its variant's Bison campaign 14/15/16).

Notes

• Idempotent: re-running on the same companies creates no duplicates (dedup catches contacts now in HubSpot).
• Clay enrichment costs credits per company + per email; keep the company cap sensible and tell the
  user roughly how many companies you're about to enrich before a large run.
• Contacts whose email can't be resolved by Clay are dropped (can't be enrolled).