leads-now-skill

A Claude skill that builds a verified contact list from free public sources — and the benchmark that keeps it honest.

Give it a description of a kind of person ("pediatric anesthesiologists within 50 miles of Myrtle Beach", "RevOps leads at Series B fintechs") and it sources them from organization websites and public registries, resolves what it can, and writes a CSV with a source and a confidence label on every field.

It also enriches a half-finished spreadsheet, grades an existing list for deliverability, and researches a shortlist for qualities a directory does not carry — such as past pediatric experience.

The rule it is built around

It never invents a contact. A guessed email that bounces costs more than a blank cell. With no known-good address for a domain it emits nothing, and says so. Anything unresolved stays empty.

That behaviour lives in deterministic code, not in prose, so it does not depend on a model choosing to be careful on a given run.

Install

Four steps, so every one is reviewable. Do not pipe it straight into your skills directory.

# 1. land it somewhere inert
mkdir -p /tmp/skill-review && cd /tmp/skill-review
curl -fsSL -O https://github.com/Scotty108/leads-now-skill/releases/download/v1.0.0/leads-now-skill.zip

# 2. verify the bytes match what was published
curl -fsSL -O https://github.com/Scotty108/leads-now-skill/releases/download/v1.0.0/leads-now-skill.zip.sha256
shasum -a 256 -c leads-now-skill.zip.sha256

# 3. look before you extract
unzip -l leads-now-skill.zip
unzip -q leads-now-skill.zip -d ./review && find ./review -type f -exec cat {} +

# 4. only then install
cp -r ./review/leads-now ~/.claude/skills/

Restart Claude Code. Confirm with "what skills do you have?" — look for leads-now.

Claude apps / Cowork: upload the zip in Customize → Skills (do not unzip it first). Cowork loads skills enabled for your claude.ai account, not your local ~/.claude/skills/.

What it does and does not do

• No network calls from any shipped code. leadkit.py imports only argparse, csv, glob, gzip, json, math, os, re, sys, unicodedata, urllib.parse. It reads local files, transforms them, writes local files.
• No credentials, no API keys, no accounts. Nothing to configure.
• No runtime instruction fetching. Every behaviour is in the files you installed, so what you review is what runs.
• All fetching goes through the host agent's own tools, under its normal permissions — never from inside a script.
• Business contact data only. No consumer or personal contact data, no scraping behind a login whose terms forbid it, no CAPTCHA or bot-detection evasion, and it does not send outreach.

Requires

Python 3.8+ and code execution. Web search and page fetching enable automatic sourcing; without them it still enriches a list you provide. Browser tools are optional and used only for sources that block plain fetching.

Repository layout

skills/leads-now/        THE SKILL — this is what ships
skill-bakeoff/           archived competitor, kept for the benchmark record
bench/                   invariant tests + the comparison protocol
scripts/build-skill-zip.sh  packages skills/leads-now/ for distribution

skill-bakeoff/finding-leads was generated by a separate skill-builder against the same brief, and the two were measured head to head for four rounds so each could absorb what the other did better. That bake-off is complete — see skill-bakeoff/ARCHIVED.md for the result and what was ported across. Only skills/leads-now/ ships; bench/LOOP.md documents the protocol that got it there.

Benchmark

python3 bench/test_invariants.py

Nine checks per skill — upload-legal frontmatter, no Claude-Code-only syntax, stdlib-only scripts, no network in scripts, the refusal behaviour, off-domain sample rejection, ambiguous-surname handling, and two measured facts the docs must not contradict.

Measured facts

Both of these are load-bearing and both were verified directly, not assumed:

• The NPI registry silently repeats past skip=1000. skip=1000 and skip=1200 return identical records with no error and no truncation flag, so a naive pager produces duplicates and a count that looks complete. Shard the query; leadkit ingest exits 4 when it detects the ceiling.
• Query the parent taxonomy, not the subspecialty. Within 50 miles of Myrtle Beach: Pediatric Anesthesiology returns 0 providers, parent Anesthesiology returns 72. The narrow query returns an empty list that looks like a correct answer.

A 50-mile radius around Myrtle Beach also spans two states — of the anesthesiology providers returned, 886 were in North Carolina. A single-state query drops every one.

License

MIT