Anonymous-visitor → outbound pipeline with suppression, revival, and ICP learner

A six-script Python harness wired as a Claude Code skill that joins a website visitor de-anonymization vendor, HubSpot CRM, the Instantly cold-email sender, Brave Search, and a Postgres prospect database into one routing pipeline. Anonymous visitors hit a webhook → intent score on the page they landed on → 5-layer suppression (existing customers, do-not-contact, recent contact, current sequence, blocked domains) → classification (e.g. agency vs end-buyer) → enrollment into the right Instantly campaign. In parallel, three sibling tools run on schedule: a 3-layer dead-deal resurrector (time decay + point-of-contact expansion + champion-follows-to-new-company tracking), a trigger prospector that monitors web signals (new hires, funding, agency searches) via Brave Search, and an ICP learning analyzer that reads approve/reject patterns from the prospect DB and recommends ICP filter changes.

The pipeline collapses four common manual loops — anonymous-traffic triage, weekly closed-lost revival, fresh-prospect harvesting, ICP refinement — into scheduled scripts with explicit handoffs. The most differentiated and most-likely-to-misfire feature is champion-follow: when a known contact leaves their old company, the resurrector picks them up at their new role and re-enters them via a context-aware sequence. Combined with 5-layer suppression running on every routed visitor, the net effect is that anonymous web traffic becomes enrolled sequences within minutes — not after the next quarterly list cleanup — and the closed-lost CRM tail goes from inert to a weekly revenue motion. Useful only if the deployer already runs this stack; not a tool to adopt the stack for.

1. Stand up the de-anonymization webhook. Run python3 rb2b_webhook_ingest.py --serve --port 4100. Edit PAGE_INTENT_SCORES in the script to match URL patterns of the deployer's site (pricing pages, demo pages, doc pages score higher than blog).
2. Switch on the full automated router. python3 rb2b_instantly_router.py --serve --port 4100 chains score → suppress → classify → enroll in one pipeline. Edit data/campaigns.json with the deployer's Instantly campaign UUIDs and AGENCY_KEYWORDS_* for their classification logic.
3. Validate the 5-layer suppression manually. Before enabling live enrollment, run python3 rb2b_suppression_pipeline.py --email user@example.com on a known existing customer to confirm all 5 layers fire.
4. Schedule weekly dead-deal revival. python3 deal_resurrector.py --top 10 --dry-run first — review drafts and loss-reason tagging. Edit LOSS_REASON_BONUS for the local CRM's close reasons. Drop --dry-run only after a clean dry-run pass; live mode sends emails.
5. Schedule weekly trigger prospecting. python3 trigger_prospector.py --days 7 --top 15 harvests fresh prospects from web signals via Brave Search. Edit SEARCH_QUERIES to match the target market.
6. Periodically run the ICP analyzer. python3 icp_learning_analyzer.py reads approve/reject patterns from the Postgres prospect DB and emits filter recommendations. Treat the output as a proposal — apply selectively, never auto-overwrite ICP config.
7. Operate the suppression layers as code. All 5 layers (existing customers, do-not-contact list, recent contact, in-current-sequence, blocked domains) are programmatic; manual review queues exist only for the dead-deal resurrector and the trigger prospector, not for the live visitor router.

Commands (verbatim):

python3 rb2b_webhook_ingest.py --serve --port 4100
python3 rb2b_suppression_pipeline.py --email user@example.com
python3 rb2b_instantly_router.py --serve --port 4100
python3 deal_resurrector.py --top 10 --dry-run
python3 trigger_prospector.py --days 7 --top 15
python3 icp_learning_analyzer.py

Data flow (verbatim from skill):

RB2B Webhook → Ingest (score) → Suppress (5 layers) → Route (classify) → Instantly
HubSpot CRM → Deal Resurrector (score + draft emails) → Review Queue
Brave Search → Trigger Prospector (score + enrich) → Outreach Queue
Prospect DB → ICP Analyzer (learn patterns) → Filter Recommendations

Install (verbatim — keep author path runnable):

git clone https://github.com/ericosiu/ai-marketing-skills.git \
 && cd ai-marketing-skills/sales-pipeline \
 && pip install -r requirements.txt \
 && cp.env.example.env \
 && cp SKILL.md <your-project>/.claude/skills/sales-pipeline.md

Required env vars (verbatim from .env.example):

HUBSPOT_API_KEY= # deal resurrector + suppression
INSTANTLY_API_KEY= # router + suppression
BRAVE_API_KEY= # trigger prospector
DATABASE_URL= # ICP analyzer (Postgres)

• Champion-follow misfires. Following a known contact to their new company is the highest-yield feature and the most likely to embarrass the sender if loss-reason tagging is sloppy (e.g. lost-to-competitor reasons routed as champion-departures). Audit loss reasons before turning revival live.
• Dry-run discipline. Deal resurrector's --dry-run flag is essential for first runs; live mode sends emails on the first script invocation. Treat the flag as a hard gate, not an option.
• Intent scoring and agency classification drift. Both are config-driven, not learned. They need manual upkeep as the market shifts — a new ICP segment can sit unrouted for weeks if URL patterns aren't refreshed.
• Heavy external dependency surface. Five paid/external services. Adopt only if the deployer already runs the stack — this is not a workflow to adopt the stack for.