List builder skill

List Builder — the meta list-building process

⚡ OPERATE VIA THE RUNBOOK

The manual phase-by-phase flow below still documents the concepts, but the OPERATIONAL path is now three one-command scripts — see RUNBOOK.md in this skill directory for the decision table and failure playbook. In short:

npx tsx scripts/make-judge.ts --spec=judge-spec.json --out=prompt.txt   # judge (mandatory blocks)
npx tsx scripts/run-lane.ts  --config=lane.json                         # full lane, resumable
npx tsx scripts/snowball.ts  --config=lane.json                         # sweep until dry
npx tsx scripts/contacts.ts  --config=lane.json                         # contacts + emails after approval

Re-running the same command resumes after ANY failure. summary.md line 1 = READY / NOT READY with the next command. Cross-run dedup via the OPTIONAL list_builder_judged_domains Postgres registry (WAL-first; scripts/registry.ts; DDL in references/registry-schema.sql; leave LIST_REGISTRY_DB_URL unset and it degrades to per-run WAL dedup). Prospeo pacing is globally rate-limited across processes (lib.ts slot reservation — never tune it).

One entry point for every list request. This skill owns the PROCESS and delegates to tool skills that own the provider gotchas. Read the referenced skill whenever you touch its provider.

Layer	Delegate
Company discovery + qualification + snowball	/list-expander (scripts in its scripts/)
Prospeo filters/syntax	/prospeo-search-api
Blitz domain→people	/blitz-list-builder (scripts/find-contacts.ts)
GetLeads counts/exports	scripts/getleads-client.ts (counts are free) — optional, skipped without a key
Email validation	provider emails from GetLeads/Blitz/Prospeo → leads-final.csv; validate with MillionVerifier (/cold-email-starter-kit) before upload
Judge tuning	/icp-prompt-builder
Local/SMB discovery	/google-maps-list-builder
List QA before delivery	/list-quality-scorecard

Scripts in THIS skill's scripts/: contacts-merge.ts (normalize/dedup contact pulls, keeps every raw provider field), enrich-domains.ts (batch Prospeo enrichment for any domain list — headcount/revenue/state/ phone; strips subdomains), classify-batch.ts (multi-segment nano classifier — one pass assigns each company to one of N segments; use for industry sweeps that feed several lanes).

Observability + multi-lane helpers (2026-07-14):

• fleet.ts — read-only status board over ~/output/list-builder/lanes/*: one row per lane (current stage, status, rows in newest stream, minutes idle, snowball in flight, READY?). Zero side effects — safe to run against live runs. npx tsx scripts/fleet.ts (add --json).
• multi-lane-classify.ts — when several lanes for one client draw from OVERLAPPING pulls, union their MERGE candidate pools by domain and label every unique company in ONE nano pass (labels = lane names, prompt built from each lane's judge ICP one-liner). Writes per-lane <client>-<lane>-candidates.csv for consumption via extra_candidates. Additive — does not change run-lane's per-lane flow. npx tsx scripts/multi-lane-classify.ts --lanes=a.json,b.json.

Requirements / env

Prerequisites: Node ≥ 18 with npx tsx (npm i -g tsx); python3 + pip install gspread only if you use the Google Sheet destination; psql on PATH only if you set LIST_REGISTRY_DB_URL. Copy the repo-root .env.example to .env and fill in the keys below — scripts walk up from skills/list-builder/scripts/ to find it.

Put these in a .env at the repo root (or ~/.env — loadEnv() reads both) or export them.

REQUIRED

Var	What for	Get one
PROSPEO_API_KEY	company + people search, the primary discovery source	sign up at https://prospeo.io
OPENAI_API_KEY	the GPT-5-nano judge that scores every candidate	sign up at https://platform.openai.com

OPTIONAL — each of these degrades gracefully: the script logs a skip line and continues.

Var	What for	Without it
OPENAI_API_KEY_NANO	a separate key/project for nano scoring	falls back to OPENAI_API_KEY
OPENAI_ICP_MODEL	override the judge model (default gpt-5-nano)	uses the default
GETLEADS_API_KEY	free contact counts + exports, first stop in Phase 4 (sign up at https://getleads.io)	contacts.ts marks GETLEADS skipped and every domain falls through to Blitz/Prospeo
BLITZ_API_KEY	domain→people fallback via /blitz-list-builder (see that skill for signup)	BLITZ stage skips, Prospeo covers those domains
EXA_API_KEY, PARALLEL_AI_API_KEY	lookalike / entity discovery inside /list-expander Phase 2 (sign up at https://exa.ai, https://parallel.ai)	those sub-sources are skipped, seeds + filters still run
LIST_REGISTRY_DB_URL	Postgres URL for cross-run dedup (DDL: references/registry-schema.sql)	WAL-only dedup: a run never re-judges itself, but earlier runs are not deduped
MILLIONVERIFIER_API_KEY	validate leads-final.csv before upload (https://millionverifier.com)	OPTIONAL for the build; REQUIRED before you send
GOOGLE_APPLICATION_CREDENTIALS	service-account JSON for the Google Sheet destination (push-sheet.py)	omit destination.sheet_id in lane.json for a CSV-only lane

psql must be on PATH for the optional registry. Nothing else is required.

Emails

GetLeads, Blitz and Prospeo each return an email + verification status per person. contacts.ts writes every contact that has one to leads-final.csv with an email_status column (verified when the provider says so, else unverified). Before you upload, run the whole file through MillionVerifier (see /cold-email-starter-kit) and keep only ok/valid results — never catch-all, never unknown.

Hard rules

1. ALL matching titles at ALL companies — never cap contacts per company.
2. Emails: validate every address before sending. A provider status of verified/VALID is a hint, not send-ready; MillionVerifier ok is the gate.
3. Keep only strict valid results. Never catch-all/unknown.
4. Save the MAXIMUM data providers return — every contact row carries raw_json; database destinations get a raw_json JSONB column.
5. A live Google Sheet progress view is the default whenever you have sheet credentials (scripts/push-sheet.py, destination.sheet_id in lane.json) — stream verified rows as they land. Without credentials the lane is CSV-only; everything else is identical.

Phase 0 — Intake (ask ALL of this up front, one message)

1. Seeds: ~10 known-good companies (client's customers are best). If none given, derive candidates and confirm.
2. ICP one-liner + explicit disqualifiers (what looks close but is NOT a fit).
3. Bands: employee range, revenue floor, founded, etc. — applied AT PULL TIME.
4. Geography.
5. Target titles/seniorities (no cap on matches).
6. Destination: database table / Google Sheet / local CSV — ask; any combo.
7. Exclusions: any CSV/table of domains to skip (optional; do NOT auto-dedupe against your sending platform — that is handled at upload time).

THE EXHAUSTIVE-SWEEP LAW (this is the point of the skill)

Recall is sacred; precision is nano's job. "I don't care how much money we spend on AI to clean these things up. I want the largest list we can possibly grab."

1. Industries: if even ONE confirmed-fit company carries an industry tag, pull EVERY company in that industry (within band + geography) and nano-score all of them. A hedge fund tagged "Financial Services" means you pull all of Financial Services in scope. Never pre-filter an industry pull by keywords.
2. Keywords: every discriminative keyword/phrase found on confirmed fits gets pulled across ALL industries (you have no clue what industries a "SaaS"-tagged company sits in). Multi-word phrases beat their parts: "medical" bad, "group" bad, "medical group" good. Drop only stopword-grade single tokens; when unsure, pull it and let nano sort.
3. The ONLY pre-filters allowed at pull time are the client's hard band (headcount/ revenue) and geography. Everything else is decided by the GPT-5 Nano judge reading the company's description/homepage.
4. Score EVERYTHING pulled. Never sample, never cap, never skip scoring to save money — nano costs ~$1-3 per 10k companies; a missed segment costs a client.
5. Snowball: every nano-confirmed fit enlarges the seed set → re-derive industries + keywords from the LARGER set → sweep the delta → repeat until a round adds <2-3% net-new qualified companies. That convergence is the only valid reason to stop.
6. The filter-precision scorecard is REPORTING (transparency for the client), never a gate that excludes a filter from being pulled.

Corollary: hand-picking a few keywords "for speed" is a bug, not a shortcut — an Investment Management industry sweep surfaced several large funds that keyword search missed entirely.

Operating rules:

• Industry trigger is LITERAL (1 confirmed fit = sweep the industry). Mega-industries (>100k companies in scope) go through the SAME Prospeo /search-company pull: pull.ts auto-shards any filter whose page-1 total_count exceeds 24k — country-wide into 50 states + DC, then a single state by headcount-band bisection — so Prospeo's ~25k-per-filter tail truncation never silently drops the remainder. snowball.ts logs [mega] <industry>: N in scope when it hands one of these to the sharder.
• Keywords: pull unless stopword-grade. Multi-word always; specific single tokens ("SaaS", "REIT", "Medicare") yes; generic ("group", "services") no; unsure → pull.
• Cost gate: run without asking below ~500k total sweep; above that, show the operator the count/cost breakdown first.

Phase 1 — Discovery (cast every net in parallel)

Run ALL applicable sources, dedup by domain (normDomain in list-expander lib.ts):

• Seed-driven core (when seeds exist — almost always): /list-expander Phases 1–3 (fingerprint → lookalikes → filter mining with precision×volume scorecard).
• Prospeo /search-company with the winning filters (450ms pacing; re-pulls free 30d).
• GetLeads count_contacts/search_contacts with company_description + technographic filters (free counts; a third corpus).
• Blitz POST /v2/search/companies (free) when company-first.
• Google Maps when local/SMB — /google-maps-list-builder (live scrape by category × metro). Category names are ground truth for local trades (e.g. 'Fence contractor'), so sweep the category rather than a keyword guess. If you keep a bulk Maps export on disk (Common Outbound Lists/Google Maps Scrape - 12M US Businesses/), grep it by category first (free) and feed the result in as extra_candidates.
• Shopify / DTC — start from the bundled Common Outbound Lists/shopify-stores-259k-active.csv.zip (259K Active stores with ≥1k estimated monthly visits; domain, merchant_name, description, company_linkedin). Unzip and pass the CSV as extra_candidates. Still run /icp-prompt-builder + live-site verify; ~11–12% of snapshot stores go dead.
• Parallel entity-search / Exa findSimilar / DiscoLike — already inside list-expander Phase 2 (see /disco-like). For local-service TAMs, ALSO run a Parallel per-metro fan-out (one objective per metro, match_limit ~40, ~$0.005/call, then keyword-filter the output — roughly 75% is padding). Parallel returns NO domains: resolve them by company-name search in Prospeo /search-company, then Exa search (needs credits).

External-source CSVs feed the lane via extra_candidates in lane.json: absolute CSV paths with a domain column; rows merge into MERGE with normDomain + registry dedup + emp-band-when-known. This is how Maps pulls, Parallel fan-outs, and ad-hoc keyword pulls enter the judged pipeline.

Post-lane EXPANSION pass for local trades (measured +56% on top of a READY lane): after a lane is READY, sweep LIVE sources and run the delta through judge→verify→registry:

1. Live Maps by QUERY (not category) — see /google-maps-list-builder; ~350 cities × trade queries, keep the website field; ~70% net-new versus category-based candidates.
2. Yellow Pages — Apify trudax/yellow-pages-us-scraper, ONE run with startUrls = yellowpages.com/search?search_terms=X&geo_location_terms=<STATE> × 50 states. Spam-dense (36% aggregator/doorway junk, 25% judge-pass) — never skip judge+verify. automation-lab actor returns NULL websites.
3. Yelp — Apify memo23/yelp-scraper (startUrls = yelp search URLs, fetchBusinessDetails) — ~90% website fill + businessOwnerName (owner first names → contacts phase); shallow yield. Both Apify actors are optional: skip them and the lane still completes. Delta pipeline: dedup vs lane WAL → score-batch.ts → verify-website.ts (--run=__abs__<dir>, stream files must be named pull-batch*-scored.csv.stream.csv) → walAppend + registry.ts sync → append tab to the destination sheet.

Phase 2 — Qualify + verify (greedy discovery is safe because of this layer)

• Tune the judge once via /icp-prompt-builder (10-company rounds until 2 clean).
• Scale-score with list-expander/scripts/score-batch.ts (GPT-5 Nano, OPENAI_API_KEY_NANO or OPENAI_API_KEY, concurrency 60, --scrape for missing descriptions; streams to .stream.csv).
• Always run list-expander/scripts/verify-website.ts — ~10% of DB-qualified companies are dead/parked ghosts; live homepage re-judge catches stale-DB false positives.

Phase 3 — Snowball until dry (the "largest possible list" guarantee)

Feed each round's confirmed fits back as new seeds → new lookalikes → new candidate filters → pull the delta → qualify. Stop when a round adds <2-3% net-new qualified companies. Record rounds + net-new counts — the convergence curve goes in the client report as proof of full-market coverage.

Phase 4 — Contacts (uncapped, cost-ordered)

For every qualified company, pull ALL people matching target titles:

1. GetLeads (free counts/exports, optional key): search_contacts/export_contacts filtered by domains (batch ≤500), job_titles/seniority. Download export CSV. No key ⇒ the stage is skipped and every domain falls to Blitz/Prospeo.
2. Blitz (optional key): /blitz-list-builder scripts/find-contacts.ts for domains GetLeads covered thinly.
3. Prospeo /search-person with company.websites {include:[...]} (LAST — most expensive; strip subdomains or the whole batch 400s).

Merge + dedup: npx tsx scripts/contacts-merge.ts --run=<slug> --csv=<file>:<provider> ... (dedupes by linkedin_url, else domain+first+last; keeps every original column in raw_json).

Phase 5 — Emails

contacts.ts step 6 writes leads-final.csv (every merged contact with a provider email, email_status = verified|unverified). Then validate:

# MillionVerifier bulk validation — see /cold-email-starter-kit for the script and key setup
# keep result == "ok" only; drop catch_all / unknown / invalid

Phase 6 — Deliver

• Chosen destination(s): a per-client database table (raw_json JSONB; RLS on if you use Supabase) and/or Sheet and/or CSV.
• The sheet, if configured, is already populated from Phase 2 onward.
• Transparency report: list-expander/scripts/report.ts + snowball convergence curve + contact/email funnel. Verified-email ceiling via contact-count.ts (free) BEFORE Phase 4 so the client sees the ceiling early.

Automation hardening (2026-07-14)

The orchestrators now self-protect against the failure classes that caused silent misses:

• Auto-shard >24k Prospeo filters (pull.ts): any single filter whose page-1 total exceeds 24k is auto-sharded — country-wide → 50 states + DC, then a single state → headcount band bisection (recursively), merged by domain. Logs sharded <label> into N sub-pulls. Inherited by run-lane PULL and every snowball pull; no more lost tails.
• PRECHECK config validation (run-lane.ts): probes Prospeo once with the emp band and once with ALL industries in one call, and rejects emp_max>100000 (unless a no-band lane). A bad band/industry now fails PRECHECK with the exact offending field and BLOCKS pull — no wasted run.
• Legacy industry map (list-expander/lib.ts LEGACY_INDUSTRY_MAP / mapIndustry): snowball translates legacy LinkedIn tags (Manufacturing → General Manufacturing, etc. — all probed valid) to the modern Prospeo enum before sweeping, so a legacy tag sweeps its modern equivalent instead of erroring the whole set. Add only names verified valid via a free count probe.
• snowball rounds now run the same REJECT_AUDIT rescue as run-lane on each round's rejects, check child exit status + output length (throw a resume message instead of silently continuing), use a two-phase swept-ledger (a filter is "done" only after a judged entry lands — a crash between pull and judge re-pulls, free within 30d; old single-phase ledgers still honored), and stream dedup round rows straight to snowball-rN.csv (flat memory).
• score-batch straggler deadline + retry: each row's classify (incl. scrape) has a 120s hard timeout (marks scored=false reason=timeout, no more 99.99%-done hangs), and one automatic retry pass re-scores all failures before reporting the final count.

Ops notes

• Run a 50–100-row test before every full run; read the first progress output and confirm real successes; surface failures immediately rather than at the end.
• Prospeo pacing is 450ms and globally enforced across processes. MillionVerifier 403s the default Python/Node User-Agent — send a browser UA.
• Reference funnel from a real pilot: 10 seeds → 30,405 pulled → 11,198 qualified → 9,247 live-verified → 1,727 in-band → ~30,755 verified-email ceiling.