From data dump to deal flow
Building a prospect list you'll actually work. Vet the source, pull the roster, research it properly, score it, and start at the top on Monday.
A late pivot. Michelle Hamilton's Build Your Own AI Chief of Staff moved to September 18, so Ivy took the August slot with the workflow she'd built the week before — and ran it live, unrehearsed, in front of everyone. Her words: "Fair warning, we pivoted this so last minute… we're going to test it live with you guys."
Everyone in this group has the same drawer: a stack of business cards, a conference export, a directory download. Two hundred names you fully intended to research. You got through the first one or two and then called whoever was easiest to reach. The gap was never finding the data. It's everything between finding it and knowing who to call first.
About that live demo
Short version: it stalled. Screen-share permissions, a surprise logout, and a list Claude never finished pulling — "as you can see, finicky, but when you can get him to work, he's great." The workflow is what you came for, so here it is properly, in the order you'd actually learn it.
What you should be able to do by Monday
Five capabilities, each one built on the last. Work them in order — skipping to the scoring without vetting the source is how people get themselves into a conversation they don't want to have.
Judge a source yourself, in about a minute
Look at any directory, roster, or member list and reach a defensible green / yellow / red — without asking anyone's permission and without guessing.
You've got it when: you can point at the specific robots.txt line or terms clause that decided your verdict, and you've logged it so you never re-read it from scratch.
Get rows out of a page, not a wall of text
Set the filters inside the platform's own interface, then have AI read only the result set that's rendered on screen — no wandering into individual records, no invented columns.
You've got it when: your row count matches the result count the page displayed. If it doesn't, something got dropped or invented, and you want to know which.
Enrich with sourced research — and let a gap stay a gap
Four separate passes: is it public, what do investor relations say, what changed recently, who leads it. Every answer carries a citation, because an uncited answer is a guess wearing a suit.
You've got it when: every figure has a source beside it and the words "not found" appear somewhere. A sheet with no gaps in it is a sheet that made things up.
Turn a list into an order of operations
Write down your own filters — size floor, geography, asset types, hard disqualifiers — then score fit and reachability and sort by the total. Your criteria are the part AI genuinely cannot supply.
You've got it when: you can name your first five calls for Monday morning, and every row you cut is still visible with the reason it was cut.
Make it repeatable so you never rebuild it
Once the sequence works for your business, save it as a skill. The next conference export or stack of business cards runs through the same machine instead of starting from a blank chat.
You've got it when: a new list is a ten-second instruction, not an afternoon — and the research runs in the background while you do something else.
Ground rules that apply to all five
- Be the human in the room. "AI cannot replace you and your business intelligence that lives rent free in your head."
- Use a paid subscription. Claude, ChatGPT, Copilot — whichever. Paid tiers protect your data in ways free tiers do not.
- Fact-check it. Watch the data. If a number matters, verify it at the source.
- Make it cite itself. The single highest-value setting in the whole workflow, and the one most people skip.
- Give it room to work. Real research on real companies takes real minutes. Kick it off and walk away — this is background work, not a party trick.
The method, in five moves
Check whether you're allowed to point AI at the page
Add /robots.txt to the domain — it's a public file that tells you who's welcome where. Then search the terms of use for six words: robot, scrape, crawl, automated, data mining, artificial intelligence. One minute per platform, and you log the answer so you're not re-reading terms every time. Do re-check periodically, though — Ivy's warning was that platforms are actively rewriting their AI policies right now. If you'd rather not read terms yourself, point Claude at the site and ask it to vet the page before you use it.
Filter in the platform, not in the prompt
Set asset type, market, seniority and service line in the platform's own interface, then let AI read the result set on screen. Live, Ivy went into the CREW member directory, picked United States → Georgia → industrial → property management, and got zero results. She backed off to Georgia + retail and got sixteen people — half of whom she already knew, Atlanta being her home chapter. That's the move: if you only sell in the South, don't pull a nationwide list.
Pull what's on screen into a spreadsheet
One prompt, in the Claude in Chrome side panel, with the source page open. Only what the page actually displays — no opening individual records, no filling in blanks. And if you pulled the list by hand instead, that's fine: the research half of this workflow works exactly the same on a list you gathered legitimately yourself.
Research in four layers — this is the part that matters
Ivy's view: anyone can find the data, so the value is in the enrichment. Ask separately about each layer: is the company public, what do investor relations say, what changed recently, and who is the leadership team. Then make it cite every answer. Without sourcing, it will fill gaps by inventing them; with sourcing, it tells you plainly that a given company is a private REIT and that its website only accounts for fifteen holdings — which is information you can act on. If your own CoStar report says twenty, you feed that back and ask it to research the gap. Ivy's own Claude is configured so that citing is not optional. If yours isn't, the agent setup guide on this site walks through forcing it, and Session 3 covers project-level instructions.
Apply your filters, score, and sort
This is where you do the work. Write down your size floor, your geography, your asset types, your revenue threshold, your hard disqualifiers — a civil engineer, a law firm, and a tech company are all looking for different things. Feed those in, score each row on fit and on how realistically you can reach the person, sort by the total, and work it top down. Ivy's pre-Vegas ICSC list was six thousand people. Nobody calls six thousand people. Keep the cut list, and set a 30- or 90-day cadence to pull the next ten forward.
Then save it as a skill
Once the sequence works for your business, at the bottom of the chat ask Claude to build a skill from what you just did, and save it. Next time — a stack of cards from ICSC Florida, a new conference export — you run the skill instead of the five prompts. Same for the tail end of the workflow: cadence trackers, email templates, an Outlook contact group for the top ten, automation through n8n or Zapier. Ivy also drops finished research into Notion as meeting cards, because a spreadsheet is not something you can read in the five minutes before a meeting.
The business-card trick. Photograph the stack, hand the photos to Claude, ask for a spreadsheet, and run the same research-and-score process on it. Three piles of cards on a home-office desk is a prospect list you haven't started yet.
Steal these.
Run them one at a time, in order, in the same Claude in Chrome side-panel conversation — each step's output is the next step's input. Swap the bracketed parts for your own criteria. Better yet: paste all six into Claude and ask it to tailor them to your target customer.
The leave-behind.
The worksheet, the prompt pack, and the source bank Ivy promised on the call — plus the deck and the whole workflow as an installable skill. Free, no gate, not gatekept: "we're not really selling anybody anything."
Prospect Build Sheet
The source-vetting log, your six filters, and a scoring table that sorts itself. Fill it in once and every prompt below has the answers it needs. CSV export and print.
Open worksheet Guide · Ten minutesHow to set up your agent
Standing instructions that force Claude to cite its sources, say "not found" instead of guessing, and never silently drop a row. Do it once, it applies everywhere.
Open the guide Prompt Pack · LibraryProspecting prompt pack
All six prompts are on this page above — the library is where every session's prompts live together, from Session 3's framework onward.
Open library Skill · InstallableProspect Build skill
The entire five-step workflow as a Claude skill — vet, pull, enrich, qualify, score. Drop in a new list and it runs the whole thing, refusing do-not-use sources and never silently dropping a row.
Download skill Slides · PowerPointThe deck
All 27 slides with speaker notes, including the source-verdict table and the four-check permission screen. PDF version if you'd rather read it on a phone.
Download deck Reference · On this pageThe source bank
Where the common CRE platforms and association directories actually land — green, yellow, and flat no — with the reason for each verdict.
Jump to guardrails Recording · VideoWatch the session
The full 45 minutes including the Q&A, the live demo and the stall that came with it. Hosted on SharePoint.
Watch recordingIvy has spent her career in sales and marketing, the last decade of it in low-voltage technology — fire alarm, access control, burglar and camera systems — which means ten years of explaining to people what technology can and cannot actually do. Today she works on remote video monitoring and analytics at EyeQ Monitoring. She built this workflow because she got tired of her own half-finished prospect spreadsheets, and she co-leads AI × CRE with Sarah Gudeman.
Licensed to look. Not licensed to take.
The part of the hour that saves somebody an uncomfortable conversation. Ivy is a territory manager, not an attorney — but she is the one who read the terms.
Your subscription buys access, not the database behind it.
Reading a result set you are entitled to see is exactly what the product is for. Systematically capturing the whole index into your own file is what these agreements are written to stop. The line is volume and intent, and the test that works is simple: am I reading, or am I building a copy of their database?
Two checks, once per platform. Add /robots.txt to the domain and look for the pages you use being disallowed, or AI agents named directly. Then search the terms for six words: robot, scrape, crawl, automated, data mining, artificial intelligence.
The one that catches people: at least one very large CRE platform revised its terms in July 2026 to bar sharing passcodes with "any third-party AI services, including, without limitation, artificial intelligence agents," and separately bars exposing the product to outside AI tools. Even logged in, even as a paying subscriber — that one is a flat no. Run the six-word search on your own subscriptions before you run any of this.
And you are identified. When you're signed in, the platform can see exactly what you're doing and who did it. That cuts both ways: don't give them a reason to ask questions, and don't assume anonymity you don't have.
Not legal advice. Sources checked August 2026 — terms change, so check again before you rely on any of this. If your firm has counsel, this is a fifteen-minute conversation worth having with them.
What do the terms say?
Search for six words: robot, scrape, crawl, automated, data mining, artificial intelligence.
What does robots.txt say?
Add /robots.txt to the domain. Look for the pages you use being disallowed, and for AI agents named directly.
Reading or harvesting?
One filtered result set is reading. Walking the whole index into a file is not. Be honest about which.
What else is on screen?
Claude sees the tab it's given. Run it in a separate browser window with only the pages you're happy for it to read. Not your bank.
Filter, then read one screen
Set the filters in the platform's own interface, get the result list on screen, and have AI read only what's rendered there. You filtered with a tool you licensed and read something you were entitled to see.
Platforms that name AI agents
Some terms now name artificial intelligence agents directly and bar sharing credentials with them. If yours does, there is no compliant version of this. Check before you assume.
Where the terms are silent
Some member directories have no automation clause at all. The constraint there isn't a contract — it's the code of conduct and your standing with people you'll see at the next chapter meeting. If you'd be uncomfortable explaining it at that meeting, that's your answer.
"But I'm a member. Doesn't that change things?"
Yes, and not in the direction you're hoping. Signing in is how you accepted the terms. Your membership buys you the right to look someone up. Almost none of these agreements grant you the right to take the whole directory.
For the record, Claude in Chrome is not blocked from logging in — signing in is permission-gated, not prohibited, and there's even a password-manager integration that fills credentials without the assistant seeing them. The constraint was never the tool. It's the agreement you accepted when you signed up.
Also worth knowing what each platform will and won't hand you. Ivy's example: on some association sites you can pull the person and their role but not their email address — for that you'd need a paid data subscription of your own. That's a licensing question, not a prompting one.
What the group asked.
Questions ran alongside the demo for most of the hour. Each one below is a real obstacle that stops people at a specific step — so the answer includes what to actually do about it.
No. A connector is convenience, not a requirement, and waiting for one is how a workflow stays theoretical. The reliable workaround is to make your CRM's import template the output format: build a sheet whose column headers exactly match the fields your CRM expects, hand it to Claude, and tell it to fill that rather than inventing its own columns. You process the whole list at once and upload in a single action.
Where a connector does exist, use it — Ivy's team runs Zoho, which connects to Claude directly. Exporting to Excel and working the data there is the same idea with more steps.
Do thisExport a blank import template out of your CRM once, and keep it as the required output shape for every list you build from now on.
Because the extension reads what is rendered on the tabs you hand it, and it cannot unsee something it should never have been shown. This is the single most common way people create a problem for themselves — not by scraping something they shouldn't, but by leaving a client portal open in the next tab.
Claude in Chrome shows you exactly which tabs it can reach. Ivy keeps one window purely for Claude to work in and does everything personal, banking very much included, somewhere else.
Do thisBefore any pull, close or move every tab that isn't the source page. Client portals and your inbox are the two that catch people.
The browser half is off the table, and that's a legitimate answer rather than an obstacle to route around. Firewalls and extension allow-lists exist for real reasons, and working around IT is a much faster way to lose access than to gain it.
But the browser was never where the value was. Steps 1 and 2 save you a few minutes of copying; steps 3 through 5 — sourced research, filtering, scoring — are where the hours actually are, and they run in an ordinary Claude chat on a list you gathered by hand.
Do thisPull the list manually, paste it in, and start at Goal 03. Then ask IT what is approved — the answer is often "the desktop app, just not the extension."
Installation is unremarkable: Chrome Web Store → Extensions → add it → sign in with your Claude account, the same flow as the Excel add-in. Check marks then show which windows it's live in.
The setting that actually matters is the approval mode. Automatic approval lets Claude check its own work against its own safety rules. Skip-all-approvals means nothing is checking anything. Manual approval makes it ask before every action, including before it enters anything sensitive.
Do thisLeave approvals on. Being asked is the entire point of having a human in the loop — the extra clicks are the product working, not the product being slow.
The site prompt library collects every session's prompts in one place. But the honest answer is that better prompts are the smaller lever. A well-instructed setup — a project or agent told to cite every claim, stay in your voice, and say "not found" rather than guess — beats a clever one-off prompt every single time, because it applies to all of them.
The improvement loop is unglamorous: run the task, notice what you had to correct, go back and fold the correction into the prompt, save that version.
Do thisKeep one running file of prompts that worked. When it stops being a file and starts being a process, ask Claude to turn it into a skill.
Shared chat links behave like any other link: assume anyone holding one can open it. Ivy flagged she'd seen something about shared Claude chats being publicly visible — she didn't have the source in hand and we haven't verified it, so treat it as a reason to review your own shared links rather than as a confirmed incident.
The broader habit is the one that applies to every tool in this workflow: know what you've made accessible, to whom, and for how long. That's a question you can answer in five minutes and can't answer retroactively.
Do thisAudit your shared chats. Anything containing a client name, pricing, or a member list should not be sitting behind a live public link.
Three things the group needs from you.
Convention happy hour
We're planning an AI × CRE happy hour on the Tuesday of CREW Convention at the Hotel Blue bar. Reserving the space carries a fee — if you or your company would like to sponsor it, tell Sarah or Ivy.
Offer to sponsor →Member spotlight
Five minutes at the top of a future session for one member to show a win or an aha moment from the past month — a tool, a prompt, a workflow that finally clicked. It's been on the agenda for a while and never used. Volunteer and it's yours.
Claim a slot →Tell us what to cover
The website is the group's resource hub, and the form fills on it feed the agenda directly. Speakers you want to hear, topics you want covered, prompts you're using, a podcast or article worth sharing — drop it in and the working group routes it.
Open the survey →Next up: Michelle Hamilton on building your own AI Chief of Staff, Friday September 18, 10:00 AM ET. The working group meets right after each session — it's casual, and they'd take the help.
Watch it back.
The full 45 minutes, including the Q&A and the live demo that decided to stall on camera.
Everything Ivy promised on the call, the prompts, the worksheet and the source bank, is already above. Trouble opening the video? Email Ivy and she will sort your access.