Lead Finder: leads with receipts

Tell Claude who you sell to. Get back the companies that fit and are buying right now, ranked, with a link behind every claim.

• Facts come with a source and an exact quote.
• Guesses are written as questions.
• Unknowns stay unknown and count neither for nor against a company.

The scores come from a small script with a fixed formula, not from the model's mood, so the same evidence always gives the same score.

Install in 30 seconds

Claude Code

/plugin marketplace add machgtm/lead-finder
/plugin install lead-finder@machgtm

Claude desktop, web and Cowork: Customize › Plugins › Add › Add marketplace › machgtm/lead-finder.

Codex: the skills follow the open Agent Skills format. Install each one from its GitHub folder:

$skill-installer install https://github.com/MachGTM/lead-finder/tree/main/skills/find-leads
$skill-installer install https://github.com/MachGTM/lead-finder/tree/main/skills/check-receipts

Then just ask:

Find 25 leads. I sell a security review to B2B SaaS companies with 20 to 200 people in Europe. Must-have: sells to enterprises. Buying signals: hiring a first security engineer, recent funding.

No API keys, no paid data, no sign-up. It runs on your own Claude.

Why it exists

AI prospecting tools are very good at sounding sure. Watch what happens to a fact on its way into a typical lead list:

What was true	What the AI tool said	What Lead Finder says
They posted three engineering jobs	"Struggling with capacity"	Hiring: 3 engineering roles, posted 2 Oct, with the link
A blog post mentions a second office "some day"	"Expanding into Europe"	Unknown: no date. Ask: "Is the London office happening this year?"
No headcount anywhere on their site	"200 employees"	Unknown: headcount. Counted neither for nor against them
They raised money in 2024	"Recently funded"	Too old to count: outside the 180-day window

Buyers notice confident guesses. A list with receipts lets you write to people about things that are actually true.

What you get

For every company:

Priority	Call now · Qualify · Nurture · Skip
Fit	each of your fit signs: met, not met or unknown, with the quote and the page
Why now	dated buying signals: hiring for sales roles, funding, a new office, a new sales leader
Must-haves	a company that clearly fails a deal-breaker (say, too big) is skipped, whatever else it scores
Open questions	what couldn't be confirmed, phrased so you can ask it
First question	one opener built only from a sourced fact

Saved as leads.csv for your CRM or sheet, and leads.md to read and share.

Hiring is checked live on seven public job boards (Greenhouse, Lever, Ashby, SmartRecruiters, Recruitee, Breezy HR and Personio), so "hiring a BD manager" comes with the post and its date.

Check any email for receipts

New in 1.1: paste a cold email and ask Claude to check it.

Check this email for receipts: ...

Claude finds every claim, looks for a public page that backs it, and turns the rest into questions or takes them out. A small script scores it against the Receipts Standard: "Receipts: 4 of 6 claims (67)". You get a table of claims and sources, and a rewritten email with footnotes. It won't invent a source to make the email pass.

Want a quick check without Claude? The receipts check does a pattern check in your browser.

How it works

flowchart LR
  A["You: who you sell to"] --> B["Scorecard<br/>fit signs + buying signals<br/>weights, must-haves"]
  B --> C["Search<br/>directories, job boards, news"]
  C --> D["Check each company<br/>its own pages, quoted"]
  C --> E["hiring.py<br/>7 public job boards"]
  D --> F["leads.json<br/>evidence only"]
  E --> F
  F --> G["score.py<br/>fixed formula"]
  G --> H["leads.csv + leads.md<br/>ranked, every claim linked"]

Claude does the reading. The scripts do the maths.

The formula

Score	How it's worked out
Fit	weight of signs met ÷ weight of signs known, 0–100. Unknown signs count neither way.
Coverage	weight of signs known ÷ all signs. Below 50%, fit is unproven: the company can only be Qualify or Skip.
Intent	each signal type counts once, at its freshest: weight × recency, summed, ÷ total signal weight. Recency: 0–30 days ×1, 31–90 ×0.6, 91–180 ×0.3, older ×0.
Priority	Call now: fit ≥ 60 and intent ≥ 30. Nurture: fit ≥ 60, no recent signal. Qualify: a signal but fit unproven. Skip: the rest, or a failed must-have.

Every rule is in skills/find-leads/reference/scorecard.md, and the script is about 300 lines of plain Python you can read in ten minutes.

Ground rules

• Public pages only. No LinkedIn scraping, nothing behind a login.
• Companies and roles, never personal emails or phone numbers.
• Everything dated, so you know how fresh it is.
• If fewer companies survive checking than you asked for, it tells you instead of padding the list.

What it connects to

There is no Lead Finder server, and nothing is sent to MachGTM. Everything runs in your own Claude session.

Connection	What for	What is sent
Claude's web search and web fetch	Finding companies and reading their own pages, careers pages and news	Your search queries and the pages Claude opens
Seven public job-board APIs (read-only, no key): Greenhouse, Lever, Ashby, SmartRecruiters, Recruitee, Breezy HR, Personio	hiring.py checks open roles and the date each was posted	A guessed board name per company (for example acme), with the user agent lead-finder/1.0 (+https://machgtm.com)
Your disk	Writing leads.json, leads.csv, leads.md, receipts.json and receipts.md	Nothing

여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗