Prompt-based prospecting: how to build a verified lead list with Claude or ChatGPT

The most common mistake I see with AI prospecting is asking ChatGPT for "50 VPs of Sales at Series B fintech companies, with emails" and trusting the neat table that comes back. It looks finished. Then the bounces start. The usual conclusion is "AI prospecting doesn't work". I think that's wrong. The prompt was fine. The setup behind it was missing.

The short answer: prompt-based prospecting means describing your ideal customer in plain language to Claude or ChatGPT, and letting a connected data layer do the searching, enrichment and verification. The model plans and reasons; it should never be the source of the contact data. The setup that works has three parts:

• An AI client you already use: Claude or ChatGPT.
• A data connector over MCP. I'd use the Bitscale MCP server, which puts 18+ contact-data providers in one email and phone waterfall behind a single connection.
• Your CRM, so the finished list lands where reps work, not in a chat window.

Some people have started calling this "vibe prospecting". The name plays on "vibe coding", which Andrej Karpathy coined in February 2025 for building software by describing what you want. The comparison is half right, and the wrong half is the one that costs you your sender reputation.

What prompt-based prospecting is (and what it isn't)

My working definition: prompt-based prospecting is list building where the instructions are written in natural language and every fact in the output comes from a data source the model called, not from the model's memory.

The second half of that sentence is the whole point. A language model is very good at turning "mid-market SaaS companies that just started hiring SDRs" into a structured search, and at deciding what to do when a search comes back thin. What it is not is a contact database. Here's the distinction I'd draw:

If you want the broader case for running GTM work from an AI assistant, I wrote about it in Claude Code for B2B sales. This post stays narrowly on building the list.

Why "just ask ChatGPT for leads" fails

The model providers are clear about this. Anthropic's own guidance on reducing hallucinations says that even the most advanced models "can sometimes generate text that is factually incorrect". Its recommended fixes are to let the model say "I don't know", to ground answers in source documents, and to restrict it to provided information rather than general knowledge. The same page ends with a note that these techniques reduce hallucinations but don't eliminate them, so critical information should always be validated.

Now apply that to prospecting. An email address is exactly the kind of arbitrary, specific fact a model can't reason its way to. It can guess the pattern ([email protected]), and that guess will often look right. Looking right is the problem: a guessed address that bounces costs you more than a blank cell, because it damages the domain you send from.

Even when the assistant searches the web instead of guessing, it finds whatever ranks: old team pages and press releases from two jobs ago. Our own MCP guide tells ChatGPT users to name Bitscale in the prompt for exactly this reason, so the assistant calls the data connector instead of answering from the web.

My take: the model should own the thinking, and a verified data layer should own the facts. Every setup decision below follows from that split.

The setup: AI client, MCP data connector, CRM

1. The AI client. Claude or ChatGPT, whichever your team already pays for. In Claude you add a custom connector once on claude.ai, and it works in Claude Desktop and the mobile app too. In ChatGPT you use Developer mode, which per the Bitscale MCP guide is available on Plus, Pro, Business, Enterprise and Edu plans but not on Free. Developer mode apps are chosen per conversation, so you switch the connector on in each new chat.

2. The data connector. This is where the contacts come from. The Model Context Protocol (MCP) is the standard that lets an assistant call a tool directly. For prospecting, the connector should give the assistant company search, people search, enrichment, verification and signals. I compared the main options in the best MCP servers for sales, so I won't repeat that here. The short version of why we built the Bitscale MCP server the way we did: a server backed by a single database can only return what that one database has indexed. A waterfall tries one provider, and if it finds no valid result, moves to the next.

3. The CRM. A list in a chat transcript is gone tomorrow. A list in HubSpot or Salesforce, deduplicated against what you already own, is something every rep can work. Bitscale's two-way sync writes back only the fields that changed, on HubSpot from the Growth plan and on Salesforce from the Booster plan.

To connect Bitscale specifically: add https://api.bitscale.ai/mcp as a connector, sign in with OAuth, pick your workspace and approve. Each person connects with their own login, so every call runs with their permissions. The MCP guide says a free account gets MCP access and 200 free credits, which is enough to run the example below on a small list.

A worked example, prompt by prompt

Below is how I'd run one list end to end. The prompts are examples I wrote for this post, not transcripts of a real session, and the ICP is an illustration only. Swap in your own.

Step 1: state the ICP and the rules before any search.

Example prompt: "Using Bitscale, I want to build a prospect list. ICP: US-based B2B software companies, 50 to 500 employees, currently hiring SDRs or BDRs. Buyer: VP or Head of Sales, or Head of Revenue Operations. Rules: only use data returned by Bitscale tools, never guess an email or phone number, and tell me when a field is empty rather than filling it. Before you run anything that spends credits, show me the estimate."

Naming Bitscale keeps the assistant on the connector. The rules are Anthropic's anti-hallucination advice turned into prospecting language: permission to say "not found", and a restriction to the data provided by the tools.

Step 2: find the accounts and check the signal.

Example prompt: "Find 40 companies that match. For each, confirm the SDR or BDR hiring signal and give me one line on why it fits. Put them in a Bitscale grid, not just in the chat."

Asking for a grid gives you a table you can open in the Bitscale app, audit and rerun. Hiring is one of Bitscale's 15 buying triggers; use whichever trigger matches your offer.

Step 3: find the buyer at each account.

Example prompt: "At each company, find the VP or Head of Sales. If there isn't one, find the Head of Revenue Operations. Flag any company where you found neither, and don't substitute a different title without asking."

If decision-maker mapping is the hard part for your market, I've written a separate guide on how to find decision makers.

Step 4: run the waterfall.

Example prompt: "Run Bitscale's email and phone waterfall on these contacts. Only keep emails that pass verification. Show me the credit estimate first."

Bitscale shows a credit estimate before it runs a column or a whole grid, and the waterfall charges only for the provider that returns a valid result. A contact with no verified email stays blank.

Step 5: dedupe and sync.

Example prompt: "Dedupe this list against our HubSpot contacts and companies. Show me what's new, what already exists and who owns it. Don't write anything to HubSpot until I approve."

Then approve, and the new records sync. The assistant did the planning, Bitscale supplied and verified the data, and a person approved the write.

Quality checks before anything goes to a sequence

This is the checklist I'd hold any AI-built list to, whoever built it:

1. Every email is verified, not inferred. Spot-check five rows: you should be able to see that the address came back from a provider and passed verification. Bitscale reports bounce rates below 0.5% on its waterfall, but check your own sample anyway.
2. No duplicates, inside the list or against the CRM. The two most common sources of embarrassing outreach are emailing the same person twice and emailing someone a colleague is already working.
3. ICP fit is explicit. Each row should carry the reason it qualified: the headcount band, the location, the signal. If the assistant can't state why a company fits, drop it.
4. Titles match what you asked for, and blanks stay blank. Watch for a "Sales Manager" where you asked for a VP, and for any value nobody returned.
5. A human approves the CRM write and the first send. Assistants are good at lists. Decisions that touch customers stay with people.

What it costs

There are two meters: your AI plan (whatever you already pay for Claude or ChatGPT; Developer mode isn't on ChatGPT Free) and your data credits. With Bitscale, those are your normal Bitscale credits. Tools called through MCP use credits at the same rate as the app, so there's no separate MCP fee. Current plans:

Pricing is per account, actions are unlimited on paid plans, and credits roll over up to 3x. On the monthly prices, that works out to about 2.3 cents per credit on Growth ($349 ÷ 15,000) and about 1.6 cents on Booster ($799 ÷ 50,000). Two things keep the bill predictable: the credit estimate before each run, and the waterfall rule that you pay only for the provider that returns a valid result. Full details are on the pricing page.

If you're weighing a prompt-first workflow against a spreadsheet-first tool, I set out the trade-offs in Clay vs Claude.

The limits, honestly

I'm bullish on this way of working, but it has real edges:

• The model still makes mistakes. Connecting a data layer removes the biggest source of invented contacts. It doesn't stop an assistant from misreading a filter or skipping a row. That's why the checklist above exists.
• Chat is a poor home for large lists. For thousands of rows, build the grid through the assistant and run it in the Bitscale app.
• ChatGPT needs a little ceremony. Developer mode, the connector switched on per conversation, and Bitscale named in the prompt.
• Data quality varies by market. That's the argument for a waterfall, and for testing on your own accounts first.
• Compliance is still yours. An AI-built list carries the same obligations as any other.

How Bitscale helps

This is the gap we built Bitscale to close: the assistant is a good planner and a poor source of truth. Bitscale is the GTM data layer that sits between the two.

• Bitscale MCP server. One connector, OAuth sign-in, about 50 tools, working in Claude, ChatGPT, Gemini CLI, Cursor, Claude Code and any remote-MCP client.
• One email and phone waterfall across 18+ contact-data providers, and 150+ data sources overall. Bitscale reports 95% email and phone coverage and bounce rates below 0.5%.
• 15 buying triggers, such as hiring, so the prompt can start from intent instead of a static filter.
• An AI research agent for the fit line on each account.
• Two-way HubSpot (Growth) and Salesforce (Booster) sync that writes back only changed fields.
• Ready-made playbooks, plus done-for-you setup and migration by Bitscale GTM engineers. GTM engineer support comes with every paid plan, with a dedicated engineer on Enterprise.
• G2: 4.8/5.

Frequently asked questions

What is prompt-based prospecting?

It is building a prospect list by describing your ideal customer in plain language to an AI assistant, which then calls connected data tools to search, enrich and verify contacts, instead of inventing them.

Can ChatGPT build a lead list on its own?

It can produce something that looks like one, but models have no verified contact database and can hallucinate. Connect a data layer such as Bitscale over MCP so every email and phone comes from a provider and is verified.

How do I build a lead list with Claude?

Add the Bitscale MCP server as a custom connector on claude.ai, sign in with OAuth, then describe your ICP and rules in a prompt. Ask for a grid, run the waterfall, dedupe against your CRM and approve the sync.

What is vibe prospecting?

It is an informal name for prompt-based prospecting, borrowed from "vibe coding". The idea is the same: describe what you want and let AI do the work. For prospecting, the facts must still come from a verified data source.

Does using Bitscale through MCP cost extra?

No separate MCP fee. Tools use your Bitscale credits at the same rate as the app, you see a credit estimate before a run, and waterfalls charge only for the provider that returns a valid result.

Do I need a paid ChatGPT plan to use Bitscale MCP?

Yes for ChatGPT. Per Bitscale's MCP guide, Developer mode is available on Plus, Pro, Business, Enterprise and Edu plans, not on Free. In Claude, you add Bitscale as a custom connector on claude.ai.

Sources

Bitscale - Bitscale MCP guide - Bitscale MCP reference - Bitscale pricing

AI models and terminology - Anthropic: Reduce hallucinations - Wikipedia: Vibe coding