Have Claude Code read every company's website to find target customers

How to do it

1. 1

   Write a screening brief from your best customers

   The brief has three parts: example customers, a yes/no question every one must pass, and types that show up in searches but aren't customers.

   Prompt

   Expand ⌄Copy

   These are our best customers: [list of websites] This is what we know about them: [call notes, CRM notes, closed deals] Read them and tell me: 1. How they describe what they do, in their own words 2. Who their customers are 3. The Google Maps categories and search terms that would find them 4. Which directories, associations or vendor lists they appear in 5. One yes/no question every one of them would answer "yes" to Where the material doesn't say, write "not found."
2. 2

   Cast wide first, then remove duplicates

   Filter only by region and size; every other criterion is a guess. Search every label your customers have used, then dedupe by domain before any step that costs money.

3. 3

   Let AI read each website and judge

   Use a script to open each site and put the text into the prompt below. Judge 10 yourself first and compare with the AI; run everything only after two rounds of 10 agree.

   Prompt

   Expand ⌄Copy

   You are screening a company for a market list. Question: [the yes/no question from step 1] Answer "no" if any of these apply: [the not-a-customer types from step 1] Input: - Name: {name} - Category: {category} - Website text: {home page plus services page, first 4,000 characters} Return JSON: - answer: yes / no / unsure - evidence: the exact website text supporting the answer, unchanged - confidence: 0 to 1 - no_reason: the first "no" condition that applies, or "none" Rules: judge only from the input; if it isn't stated, answer "unsure." Never judge from the company name alone.
4. 4

   Spot-check before handing over the list

   Sample 10: 3 random, the 2 the AI was most and 2 least sure about, 2 with no website and 1 that was excluded. Open the sites to check; if any are wrong, fix the rules and sample again.