Find businesses with weak digital presence — before your competitors do.

Quick Start

Three levels. Do Level 1 today; add 2 and 3 whenever you want. Each builds on the last.

Level 1 — set up the Project (do this first)

The 4 words people mix up

Claude has a few similar-sounding features. Here's the plain-English version so you know which is which:

Project

A folder for a body of work. Its instructions apply to every chat you start inside it. Best home for this workflow — set it up once, every batch inherits the rules. Free OK

Project instructions (custom instructions)

The block of text you paste into a Project that tells Claude how to behave there. This is where our System Prompt goes.

Skill

A packaged, reusable capability (a file Claude loads when the task matches). Same brains as the Project instructions, but portable — upload once and it's available in any chat. Optional alternative to Level 1. Free OK → Install as a Skill

Connector (MCP)

A plug-in that gives Claude a new tool — like a better search engine (Exa) or live Google Maps data. You add it once in Settings. → Better Search · Exa

Rule of thumb: Project instructions or a Skill = the brains (what to do). Connectors = the tools (what Claude can reach). You want one of the first and, ideally, the Exa connector.

Better search with Exa recommended

Claude's built-in web search is fine, but it's shallow for lead research — it misses smaller businesses and thin listings, which are exactly the ones you want. Exa is a search engine built for AI. Adding it as a connector makes Claude's research noticeably deeper and more accurate. It's the single highest-value upgrade here, and the free tier needs no credit card.

Add it in the Claude app (easiest)

https://mcp.exa.ai/mcp

Plan note

Custom connectors work on Free (limited to one custom connector — make it Exa), Pro, Max, Team, and Enterprise. If you only add one connector ever, this is the one.

Using Claude Code instead? (technical path)

If you run Claude in the terminal, skip the UI and add Exa with one command:

claude mcp add --transport http exa https://mcp.exa.ai/mcp

Check it registered with claude mcp list. Add --scope user to make it available in every project.

Install it as a Skill (optional)

A Skill is the same prospecting brain as the Project instructions, but packaged as a file Claude loads automatically when you ask it to find leads. Two reasons to use one instead of Project instructions: it works in any chat (not just one project), and it's a single download — nothing to copy-paste. Pick one approach; you don't need both.

The .zip contains prospect-leads/SKILL.md — the format Claude expects for upload.

mkdir -p ~/.claude/skills/prospect-leads
# then save the downloaded file as:
#   ~/.claude/skills/prospect-leads/SKILL.md

The System Prompt

This is the brain, in plain text. Paste it into your Claude Project's instructions once (Level 1). It defines the columns, the research method, the scoring, and the honesty rules. (Prefer a file? Use the Skill instead — same content.)

You are a lead-prospecting analyst for M2 Marketing, a media-marketing agency. Your job: given a business NICHE and a geographic REGION, find real local businesses and assess how weak their digital presence is — because a weak presence is a strong sales opportunity for our services.

OUTPUT — one row per business, these exact 21 columns, in this order:
#, Business Name, Google Maps Link, City, Category, Address, Phone, Website, Owner / Manager, Role, Google Rating, Review Count, Facebook, Instagram, LinkedIn, Social Activity & Content Quality, Website Quality, Phone Reliability, Opportunity Notes, Score, Confidence / Source

HOW TO FIND THEM:
- Use web search. If an Exa search tool is available, prefer it — it's deeper. Run several queries per batch, e.g.:
  "{niche} in {city}, {state}", "best {niche} near {city}", "{niche} {city} site:facebook.com", "{niche} {city} yelp".
- Real businesses only. NEVER invent a business, phone number, address, or rating.
- If you cannot verify a field, write "unknown" and lower that row's Confidence. Guessing is worse than "unknown."

FOR EACH BUSINESS, ASSESS:
- Website: Does one exist? Is it a DIY builder (Wix / Squarespace / GoDaddy / Weebly / Wordpress.com), outdated, or missing an obvious contact form? "No website" or "DIY site" = big opportunity. Rate Website Quality: None / Weak / OK / Strong.
- Social: Find Facebook, Instagram, and LinkedIn if they exist. Judge Social Activity & Content Quality only from what's publicly visible — recency of last post, follower count, whether content looks professional. You CANNOT log in; say so and stay honest. Use: None / Inactive / Sporadic / Active.
- Reviews: Google Rating and Review Count if visible. Few reviews (<10) or low rating (<3.5) = opportunity. If not visible, mark "unknown."
- Owner / decision-maker: Check the About page or LinkedIn for a name and Role. Best-effort; "unknown" is fine.
- Phone Reliability: Note if the number looks valid and appears consistently across sources, or if it's questionable. (e.g. "verified on site + GMB", "only on one aggregator — verify").

SCORE each lead 0–100 by summing every signal it triggers:
  no website +20 | no Facebook AND no Instagram +15 | website but no contact form +15 |
  DIY website builder +10 | fewer than 10 reviews (or none) +10 | rating below 3.5 (or none) +10 |
  has social but clearly inactive/low-quality +10 | phone but no email +5 | classic service business
  (plumber, electrician, roofer, HVAC, landscaper, contractor, cleaner, etc.) +5. Cap at 100.
  Higher score = more gaps = easier, more valuable pitch.

OPPORTUNITY NOTES = the one-line pitch angle: the specific gap you'd lead with on a call.
  e.g. "No website, only a barely-active FB page — quick win with a landing page + content plan."

CONFIDENCE / SOURCE = where the data came from + how sure you are (High / Med / Low).
  ALWAYS fill this. It tells the caller what to verify before dialing.

DELIVER:
- A Markdown table with all 21 columns, sorted by Score (highest first).
- THEN offer the same data as a downloadable CSV file.
- THEN a 2-line summary: how many are hot (Score >= 70), and the top 3 to call first and why.

TONE: You are a sharp research analyst, not a salesperson. Be accurate, flag uncertainty, never pad the list with weak or unverifiable entries just to hit a number.

Run a batch

Start a new chat inside your project (or any chat if you installed the Skill) and paste this, filling in the brackets. One message per batch.

Find [30] [NICHE] in [CITY, STATE — or a list of towns].

Follow my instructions exactly: real businesses only, all 21 columns,
score each 0–100 by digital-presence gaps, sort hottest first, flag Confidence
on every row, then give me a downloadable CSV and the top 3 to call first.

Extra filters (optional): [e.g. "skip national chains / franchises",
"focus on ones with no website or a DIY site", "Spanish-speaking owners OK",
"within 20 miles of downtown"]

Getting the best results

Handy follow-ups (same chat)

• "Deep-dive rows 1–10 and update their Social Activity + Owner columns."
• "Drop anything with a Score under 40 and re-sort."
• "Write me a 3-sentence cold-call opener for the #1 lead using its Opportunity Notes."
• "Re-export just the hot leads (Score ≥ 70) as CSV."

The 21 columns

Your exact schema, plus how reliable each field is when researched from the public web. The reliability pill is what to trust vs. what to verify before you call.

Solid Claude gets this right most of the time   Best-effort often good, sometimes stale/partial — verify   Judgment a qualitative read of public signals, not ground truth

How leads are scored

Each business earns points for every weakness it has. Add them up (capped at 100). More gaps = higher score = easier, higher-value pitch. A 15-review plumber with no website and a dead Facebook is a layup; a slick brand with 400 reviews is not your lead.

Reading the score

This rubric is the exact one Flow Systems' internal prospector uses. It's a starting point — once you're tracking outcomes (next tab), you can tune the weights to what actually books calls for you.

Turn the sheet into a feedback loop

Your original ask: "figure out which factors actually correlate with successful appointments over time." Here's the low-effort way to get that — no data science, just disciplined logging plus one prompt.

Step 1 — add two columns to your Sheet

• Status — New · Contacted · Booked · Won · Lost
• Outcome Notes — a few words on what happened ("booked — loved the no-website angle", "wrong number", "already has an agency").

Update them as you work the list. That's the only ongoing discipline required.

Step 2 — after ~30–50 logged leads, ask Claude to find the pattern

Attached is my prospecting sheet. Each row is a lead, with its digital-presence
signals and Score, plus a Status column (New / Contacted / Booked / Won / Lost)
and Outcome Notes.

Analyze it:
1. Which columns or signals most correlate with leads that reached Booked or Won?
   (e.g. does "no website" actually convert better than "low rating"?)
2. Which signals look like noise — no relationship to outcomes?
3. Given that, how should I re-weight my scoring, and what should I prioritize
   when I pull the next batch?
4. Flag anything I don't yet have enough data on to judge.

Be honest about sample size — don't over-read a small number of Bookings.

Limits & honesty

Straight talk so nobody's surprised. The system is built to handle these openly rather than hide them.

What Claude genuinely can't do

• It can't log into Instagram, Facebook, or LinkedIn. Follower counts and "how active are they" come from what's publicly visible in search results — a directional read, not an analytics dashboard.
• It can't scrape Google Maps live (unless you wire up the Maps connector — see Go Further). By default, ratings and review counts come from search results, which can lag reality.
• It won't always find the owner's name. Small businesses often don't publish it. "unknown" is an honest answer, not a failure.
• It shouldn't invent to fill a quota. The instructions tell it to return fewer, real leads rather than pad the list. If it returns 22 solid ones instead of 30, that's the feature working.

The discipline that makes it trustworthy

A note on doing this ethically

This uses only public business information — the same stuff a prospect could Google about themselves. It's research, not surveillance. Keep it to business profiles and public pages, and you're on solid ground.

Go further

Two optional upgrades for when you want more accuracy or more scale. Neither is required to start.

A) Accurate ratings with Google Maps needs a card

The one thing web search gets wrong is exact Google star ratings and review counts. If those matter, connect the Google Maps data directly. Honest heads-up: this needs a Google Cloud account with a credit card on file — there's a generous free monthly allowance, but Google requires billing to be set up.

claude mcp add maps --transport http \
  https://mapstools.googleapis.com/mcp \
  --header "X-Goog-Api-Key: YOUR_MAPS_API_KEY"

B) The full tool (database + dashboard)

The Project + Sheet setup is perfect for one or a few people running batches by hand. If M2 scales this into a real pipeline, there's a heavier version — the actual tool Flow Systems runs — that Cam can help you stand up.

The full tool needs a Cloudflare account, a short deploy, and a couple of API keys — a developer setup, not paste-and-go. Worth it when you're pulling leads constantly. Until then, this page does 90% of the job for ~0% of the hassle. Cam runs the full version internally and can walk your team through their own copy when you're ready.