How to Use Claude Code for Sales Prospecting

Claude Code for sales means running prospecting from one terminal: you describe your ideal customer in plain language, Claude Code calls a B2B data source to build the list, applies yes/no qualification rules, finds the buyers, removes accounts already in your CRM and hands a clean file to your sequencer. The output is a list you can load, not a paragraph of suggestions.

The work it replaces is the part of the week that is not selling. In the Salesforce State of Sales report (fifth edition, published December 2022), reps said they spent 28% of their week actually selling. List building, research and CRM entry fill much of the rest.

A model with web search alone is a poor list builder. It surfaces a handful of well-known examples, not every account that fits a filter, and it cannot check tenure, headcount bands or your CRM. This guide shows seven steps in plain language, a worked example with a real count, the metrics to track and the mistakes to avoid. It is written for SDRs, AEs, founder-led sellers and the sales leaders who run them.

Quick answer: To use Claude Code for sales prospecting, connect a B2B data layer such as the Landbase MCP or CLI, describe your ICP in plain language, build the account list, qualify it with yes/no questions, find decision-makers at the accounts that pass, suppress CRM matches, then export to your sequencer. Review a sample before every send.

Before you start: what you need

Most bad lists come from a fuzzy brief, not a bad tool. Have these ready before the first prompt.

• Claude Code installed and a project folder for each prospecting campaign, with a short CLAUDE.md that states your offer, ICP and exclusions.
• The CLI or MCP server connected and authenticated (Step 1).
• A written ICP: industry, headcount band, geography, one or two technologies or signals, and the two or three titles you sell to. See the ICP framework if you do not have one.
• A connected CRM (HubSpot or Salesforce) with Do-not-contact and Customers rules approved, or a recent export of accounts and contacts to dedupe against.
• A decision on who reviews the list before it is sequenced, and how many sample rows they check (20 is a workable minimum).

Prospecting requests for SDRs and founder-led sellers

Good prospecting requests name every filter and the output you want. These are the shapes that work, and what the workflow does with each.

• Title + segment (SDR): ask for VPs of Sales at Texas B2B SaaS companies with 50-200 employees. The search builds the company pool on location and size, then a person search pulls the sales leaders at those accounts.
• Technology (SDR): ask for Heads of RevOps at US Salesforce users with 200-1,000 employees. Technology data is not recorded for every company, so expect a smaller list than the size filter alone suggests.
• Hiring signal (SDR): ask for logistics companies currently hiring SDRs. Job-posting counts are company-wide, so qualify for the sales team specifically.
• Funding + founders (founder-led sales): ask for founders of Seed or Series A AI startups in NYC. Company filters come first, then founder and CEO profiles at each account. The funding field holds the latest recorded round only.
• Functional buyer (founder-led sales): ask for RevOps leaders at fintech companies with 100-500 employees. Useful when the buyer is a function rather than a title.
• Lookalikes (founder-led sales): ask for companies similar to three of your best customers. Output quality follows the seeds you choose.
• CRM file (any seller): ask Claude Code to clean a CRM export. It uploads the file, matches each row, flags bad domains, merges duplicates and returns a cleaned version to review.

How to use Claude Code for sales prospecting: step by step

Step 1: Connect Landbase to Claude Code

Install the CLI using the install and connect guide, sign in through the browser and let the built-in health check confirm that setup, sign-in and connectivity work. If you prefer MCP, Claude Code can add remote MCP servers, as described in the Claude Code MCP docs. MCP is an open-source standard for connecting AI applications to external systems. Done when a test search returns a dataset. Cost: none.

Step 2: Describe the ICP in plain language

Write the ICP as one sentence with a value for every filter: industry, headcount band, location, and any technology or signal. Vague words ("tech companies", "mid-market") produce vague lists, so replace each with a number or a named category.

1. Weak: "Find SaaS companies in Texas."
2. Better: "B2B SaaS companies headquartered in Texas with 50-200 employees."
3. Best: add the exclusion, e.g. "exclude agencies and consultancies."

Save the sentence in the campaign's CLAUDE.md so every rerun uses the same definition. Cost: none.

Step 3: Build the account list

Give Claude Code the ICP sentence from Step 2 and ask it to build the account list. The reply comes back as a dataset, with the filters it applied and a suggested qualification question. Narrow in the same session (for example, only companies that use HubSpot) instead of starting over. For lookalikes, seed with specific customer domains. Ask for a CSV when you want the file. Decision rule: under 50 accounts, loosen one filter; over 2,000, add one. Default searches are free. For signals that need job-posting text or exact phrase matches, research mode drafts a query and quotes a credit estimate before anything runs; for signal-led lists see the buying signals playbook.

Step 4: Qualify with yes/no criteria

Filters get you close; qualification removes the near misses. Turn each fit rule into a question with a yes or no answer, such as whether the company sells subscription software to businesses and whether it is an independent company rather than an agency. Tell Claude Code which fields the model may read, such as name, website, industry and description.

Qualify a 25-row sample first and read the answers before the full run. If more than 1 in 5 answers looks wrong, rewrite the question. Cost: qualification spends credits. Claude Code shows a quote before the run starts; approve it only after the sample looks right.

Step 5: Find the decision-makers

Ask Claude Code to search for people inside the qualified accounts, for example VPs or Heads of Sales at the companies that passed Step 4.

Ask for two or three titles per account, ranked by seniority, plus LinkedIn URLs. For founder-led sales, ask for founders and CEOs directly. Map one economic buyer and one user per account; the buying committee guide covers larger deals. Finding emails and phone numbers spends credits, so do it after Step 6.

Step 6: Clean and dedupe against the CRM

Remove people and accounts you already own before you pay to enrich them. With a connected CRM:

1. Ask Claude Code to check the list against your connected HubSpot or Salesforce account.
2. Suppression drops people on the Do-not-contact rule and companies on the Customers rule. It does not check every CRM record.
3. For a CRM export file, Claude Code uploads and matches it, then flags bad domains and merges duplicates by domain and email.

Threshold: if more than 10% of the list matches open opportunities or customers, tighten the ICP or the exclusions. Suppression and matching are not listed among the steps that spend credits.

Step 7: Hand off to sequencing and set a cadence

Export the final list as a CSV and load it into your sequencer, or let Claude Code finish the job: it can write a message sequence for each row (charged per row that gets one), create the records in your CRM, or create an Instantly campaign (beta). Keep one owner per campaign who reviews 20 rows and the first three generated emails before launch. Rerun the saved search monthly, or whenever the ICP or the offer changes, and suppress anything already sequenced.

Worked example: Texas B2B SaaS sales leaders

The account count below is from a free default Landbase search run on November 3, 2026. The later steps describe what to do next; no qualification or enrichment was run.

• Step 2: the ICP sentence was "B2B SaaS companies headquartered in Texas with 50-200 employees."
• Step 3: the search found 34,205 Texas companies in the 51-200 employee band, then narrowed to 418 SaaS candidates using a "saas" phrase match plus a semantic match on "B2B software-as-a-service company". It held back an industry filter because confirmed SaaS examples sat under two different industry labels.
• Step 4: the search returned a ready qualification question ("Is this company likely a B2B software-as-a-service company…?") to run on the 418, starting with a 25-row slice.
• Step 5: for the accounts that pass, request VP of Sales, Head of Sales and CRO, aiming for two names per account.
• Steps 6-7: suppress customers and do-not-contact records, enrich emails for what remains, then export to the sequencer.

The useful detail is the gap between 34,205 and 418: the size and state filters alone would have produced a list that is mostly not SaaS. Qualification then removes the remaining false positives.

The honest limits of Claude Code for sales prospecting

• Seeds drive lookalikes. Seeding with your largest logos returns more large logos. Pick 5-10 seed customers that share the trait you sell to.
• Fields have definitions. Funding is the latest recorded round only, job-post counts are company-wide, and technology data is not recorded for every company.
• Industry labels are loose. As the worked example shows, SaaS companies sit under several industry labels, so qualification questions do more work than industry filters.
• Past-employer searches are weaker than current-company filters. Verify them by hand.
• The web still wins on stories. Fresh announcements, case studies and news come from search. A good pattern is web search to find candidates and structured data to verify and enrich them.

How to measure whether it worked

• Sample precision: share of 20 reviewed rows that fit every ICP rule. Hold the list until it reaches 80%.
• Contacts per qualified account: target two or more decision-makers per account before sequencing.
• CRM overlap: share of rows removed by suppression. Rising overlap means the ICP is pointing at accounts you already work.
• Positive reply rate and meetings per 100 contacts: compare each Claude Code list against your last manually built list for the same segment.

Review precision on every new list, and reply and meeting rates after each campaign's first two weeks.

Common Claude Code sales prospecting mistakes

Trusting web search for the list itself

A model with search returns a few well-known examples and sounds sure about them. It cannot filter on headcount bands, tenure or your CRM. Use the web for context and stories, and a data layer for the list.

Skipping the sample review

Even a list that is 9 in 10 correct carries 50 wrong rows per 500. Read 20 before every send and rewrite any qualification question that fails more than 1 in 5.

Enriching before deduping

Contact enrichment spends credits per row. Run CRM suppression first so you do not pay for customers and do-not-contact records.

Filtering on industry labels alone

Industry fields split one market across several labels. Use phrase and semantic matches to find the pool, then a yes/no question to confirm fit.

How to do this in Landbase

Landbase is the data layer in these steps. Describe the ICP in plain language and get a dataset of companies and contacts from 40M+ companies and 800M+ contacts, then enrich with 1,500+ fields, qualify with yes/no prompts, suppress CRM matches and export, all from Claude Code or Codex through the CLI or MCP. For other tools that work inside Claude Code, see the GTM tools roundup; for the RevOps side, see how to automate RevOps workflows with Claude Code.

Frequently Asked Questions

Can Claude Code do sales prospecting on its own?

Only partly. Claude Code can plan the search, write qualification rules, clean files and draft emails, but with web search alone it returns a few examples rather than a full list. Connect a B2B data source through MCP or a CLI so the list comes from filtered records, then let Claude Code run the qualification, dedupe and export steps.

Do SDRs need to code to use Claude Code for sales?

No. Every step in this guide starts as a plain-language request, and Claude Code writes and runs the commands. Knowing what a dataset id is and how to read a CSV helps. Most of the skill is writing a precise ICP sentence and clear yes/no qualification questions.

Which Claude Code prospecting steps cost credits?

In the CLI, default searches are free. Qualification, contact enrichment for emails and phones, research-mode SQL queries and generated message sequences spend credits. Qualification and research mode quote an estimate before they run, so approve the quote only after a sample looks right.

How is Claude Code different from an AI SDR tool?

An AI SDR tool packages list building, writing and sending in one product with fixed steps. Claude Code is a general agent: you choose the data source, the rules and the sequencer, and you can inspect every command it runs. That takes more setup but makes each step reviewable and easy to change.