/list-expander

name: list-expander
description: Expand a hard-to-build niche list from ~10 known-good seed companies into a full qualified TAM. Seed fingerprinting (how do good-fit companies ACTUALLY show up in the company database's filters) → lookalike generation (Prospeo company_lookalike, Exa findSimilar, Parallel.ai entity search) → filter mining with a precision/volume scorecard → wide pull with auto-sharding → cheap-AI qualification → live-website verification → client transparency report. Use when a vertical list feels "too small" (medical groups, hedge funds, Medicare brokerages), when known good-fit companies don't show up in keyword searches, or when someone says "expand this list", "the TAM should be bigger", "find more companies like these".

List Expander — seed companies → lookalikes → mined filters → qualified TAM

The problem this solves

Niche lists come out tiny because we search databases with narrow explicit keywords, but real good-fit companies (e.g. Atlantic Medical Group, Hackensack Meridian) often don't carry that keyword in their database record. Instead of guessing keywords top-down, this skill works bottom-up: take companies you KNOW fit, discover how the database actually tags them, expand via lookalike engines, and only then derive the wide-net filters — with measured precision per filter — before pulling and AI-qualifying at scale.

Pipeline (5 phases)

All scripts live in `scripts/`, run with `npx tsx`, and are dependency-free (node ≥ 18 native fetch). Every script takes `--help`. Keys load from the repo-root `.env` (see `.env.example`), falling back to `~/.env`. Artifacts land in `~/output/list-expander/{run}/`.

Phase 1  fingerprint.ts    seeds → how each shows up in Prospeo + its live homepage
Phase 2  lookalikes.ts     seeds → candidates via Prospeo lookalike + Exa + Parallel
                           (qualify candidates → lookalikes-confirmed.csv)
Phase 3  mine-filters.ts   confirmed fits → candidate filters → scorecard (volume × precision)
Phase 4  pull.ts           winning filters → wide pull, auto-shard, dedup, exclusions
         score-batch.ts    scale AI qualification (gpt-5-nano by default)
         verify-website.ts second pass on QUALIFIED rows: live homepage fetch →
                           dead / suspended-parked / live; live sites re-judged on
                           their CURRENT content (catches stale-DB ghosts). Always run —
                           DB descriptions happily qualify dead companies otherwise.
Phase 5  contact-count.ts  verified-email TAM ceiling (free Prospeo count trick)
         report.ts         single-file HTML transparency report for the client

Phase 0 — inputs

• ~10 seed companies known for sure to fit (domains).
• 1–2 sentence ICP description ("multi-site medical/specialty groups in the US, ≥$3M revenue").
• Optional: exclusion CSV (companies already in campaigns).

Phase 1 — fingerprint the seeds

npx tsx scripts/fingerprint.ts --domains="a.com,b.com,..." --run=<slug>

Prints coverage (which seeds Prospeo is missing — **that gap IS the under-count story for the client**), plus industry and keyword-tag frequency tables. Writes `fingerprint.json/csv`.

Phase 2 — generate lookalikes

npx tsx scripts/lookalikes.ts --domains="<seeds>" --run=<slug> \
  --objective="<NL description of the COMPANY TYPE (not your product!)>" \
  --country="United States #US" --pages=3

• **Prospeo `company_lookalike`** (required lane): `{"domain": "<seed>"}` — one call per seed,

single domain only (arrays 400). Composable with location/headcount/keyword filters. One large health system seed returned 5,727 lookalikes.

• **Exa findSimilar** (optional lane): content similarity, and it returns homepage text in the

same call — free evidence for qualification.

• **Parallel.ai entity-search** (optional lane): ~$0.005/req, pads to `match_limit` with junk —

always qualify before trusting; returns LinkedIn URLs, not domains.

Optional lanes whose key is unset are **skipped with a log line**, so the run still completes on Prospeo alone. Then qualify candidates (Claude sub-agents for a small set, or `score-batch.ts` with a draft prompt) → write `lookalikes-confirmed.csv`. Target 50–150 confirmed.

⚠️ EXHAUSTIVE-SWEEP DEFAULT

Phase 3's scorecard is for TRANSPARENCY, not selection. The pull in Phase 4 must include: (a) EVERY industry carried by ≥1 confirmed fit — whole industry, headcount band + geo only, no keyword narrowing; (b) EVERY discriminative keyword from confirmed fits across ALL industries. Score everything with the cheap model. Only band + geography are legal pre-filters. Snowball until net-new drops under 2–3%. Never trim the sweep to save AI cost — recall is the product. Pull keywords unless they are stopword-grade.

Phase 3 — mine + score filters

npx tsx scripts/mine-filters.ts --csv=<confirmed.csv> --run=<slug> --propose --country="United States #US"
# Claude reviews/edits {run}/candidates.json: prune generic n-grams, add synonym keywords
# (the "every medical group contains 'group'" trap — kill terms that are frequent but not discriminative)
npx tsx scripts/mine-filters.ts --run=<slug> --scorecard
# score the 25-company samples:
for f in ~/output/list-expander/<slug>/samples/*.csv; do
  npx tsx scripts/score-batch.ts --csv=$f --prompt-file=<icp-prompt.txt> --out=${f%.csv}-scored.csv; done
mkdir -p ~/output/list-expander/<slug>/samples-scored && mv ~/output/list-expander/<slug>/samples/*-scored.csv $_
npx tsx scripts/mine-filters.ts --run=<slug> --scorecard --precision-from=~/output/list-expander/<slug>/samples-scored

`--propose` fingerprints each confirmed company against Prospeo **and fetches its live homepage**, then mines 2–3-grams across that combined text. The homepage is the evidence source that matters: it says what a company calls itself *today*, which is exactly what a keyword filter has to match. Pass `--no-scrape` to mine from Prospeo descriptions on