Targeted Search

A thin stage wrapper around the list-builder engine in engine/list-builder/. The engine owns the process; this skill owns getting the run's ICP into it and its result back out.

Read engine/list-builder/RUNBOOK.md before running anything. Run the commands there; do not improvise. engine/list-builder/SKILL.md documents every phase and env var.

Pipeline contract

Stage 7 of 20. Reads 01-seeds.csv, 02-client-profile.yaml, 03-lookalikes.csv, and the company domains in 06-audience-qualified.csv. Writes 07-judge-spec.json, 07-lane.json, 07-accounts.csv.

Files live in runs/<run>/. Column definitions: /prospecting-system.

Steps

1. Build 07-lane.json from the profile (shape documented at the top of engine/list-builder/scripts/run-lane.ts):
   • seeds ← 01-seeds.csv
   • emp_min / emp_max, states, industries ← icp_hard_filters. Industry names must be valid Prospeo names (skills/icp-definition/references/prospeo-industries.md).
   • keywords ← the discriminating phrases from 01-customer-patterns.md
   • seniorities ← the profile's titles
   • extra_candidates ← absolute paths to 03-lookalikes.csv and a domain-column CSV of the companies behind 06-audience-qualified.csv. This is how both lanes land in one judged list.
2. Judge prompt — reuse the ICP and disqualifiers the human approved in stage 6. Write 07-judge-spec.json and generate the prompt with make-judge.ts. Never hand-write it; the template's mandatory blocks prevent known false-negative classes.
3. Test, then run. A 50–100 row test first. Then:

   npx tsx engine/list-builder/scripts/make-judge.ts --spec=<run>/07-judge-spec.json --out=<run>/07-judge-prompt.txt
   npx tsx engine/list-builder/scripts/run-lane.ts   --config=<run>/07-lane.json --run-dir=<abs path to run>/07-lane

   --run-dir keeps the lane's artifacts inside the run folder (default is ~/output/list-builder/lanes/). prompt in 07-lane.json must be the absolute path to 07-judge-prompt.txt. Re-running the same command resumes after any failure.
4. Read summary.md in the lane dir. Line 1 is # READY or # NOT READY with the next command. Only READY moves on. The failure playbook is in the RUNBOOK.
5. Copy the final qualified companies to 07-accounts.csv (domain required, keep every other column) and record rows pulled → rows qualified in state.json.

Scope for this system

The engine is built to find the largest list possible (snowball.ts sweeps until a round adds under 3% net-new). For a 500-lead run that is usually more than needed:

• Skip snowball.ts unless the lane finishes READY with too few accounts for the target.
• The engine's own cost gate asks before sweeps above ~500k companies. This system's spend cap is stricter and wins: estimate the pull (free Prospeo counts), add it to spent_usd, stop if it passes the cap.
• Judging is cheap (roughly $1–3 per 10k companies on the nano judge) — do not sample or skip scoring to save money. Save money by narrowing the pull with the hard band and geography.

Do not

• Rename or move anything under engine/ — the scripts import each other by relative path.
• Hand-edit stream CSVs or files inside a lane run dir.
• Pull contacts here. That is stage 8, after the human approves 07-accounts.csv.