How to Use Claude Code for Recruiting

Claude Code for recruiting means running candidate sourcing from one terminal. You describe the candidate profile in plain language, Claude Code queries a people-data source, narrows the pool with yes/no criteria, checks it against your ATS and hands a shortlist to outreach. The output is a file of real profiles you can review, not a list of boolean strings to paste somewhere else.

Most talent teams are still early. In LinkedIn's Future of Recruiting 2025 report, 37% of organizations said they were integrating or experimenting with generative AI in hiring, and teams doing so reported saving about 20% of their work week. The same report found that companies running the most skills-based searches were 12% more likely to make a quality hire.

A chat model with web search can write a sourcing string, but it cannot return every senior engineer who fits a filter. This guide covers eight steps in plain language, a worked example built on a live search, which steps cost credits, the metrics to track, the honest limits and the privacy rules. It is written for in-house recruiters, talent sourcers and recruiting-agency teams.

Quick answer: To use Claude Code for recruiting, connect a people-data layer such as the Landbase CLI or MCP, write the candidate profile in plain language, build a candidate pool, narrow it with free filters and yes/no qualification, map target companies, remove people already in your ATS, enrich contact details for the shortlist only, then review every profile before outreach.

Before you start: what you need

A sourcing run is only as good as the intake. Have these ready before the first search.

• Claude Code installed, with one project folder per open requisition and a short notes file holding the role brief, exclusions and privacy rules.
• Your people-data layer connected and signed in (Step 1).
• A role brief from the hiring manager intake: title, level, must-have skills, location or remote policy, and the companies or company types the best hires come from.
• An export of your ATS candidates (name, LinkedIn URL, email, status, last contacted date) for deduplication. Greenhouse, Lever and Ashby all export to CSV.
• For agencies: a list of client companies whose employees are off limits, so they are excluded before anyone is contacted.
• A named reviewer and a sample size for profile checks before outreach (20 profiles per req is a workable minimum).

How to use Claude Code for recruiting: step by step

Step 1: Connect Landbase to Claude Code

Install the CLI with the install and connect guide, sign in, and run the built-in health check to confirm the connection works. If you prefer MCP, Claude Code can add a remote MCP server instead, per the Claude Code MCP docs. The setup is the same one used for Claude Code sales prospecting. You are done when a test search returns a dataset. Cost: none.

Step 2: Write the candidate profile in plain language

Turn the intake into six attributes, and mark each one must-have or nice-to-have.

• Title and function: "senior software engineer, backend".
• Seniority: senior or staff individual contributor, not engineering manager.
• Skills: Python, distributed systems, payments.
• Location: New York, or within commuting range of the office.
• Company type and stage: fintech, 50-500 employees, venture-backed.
• Tenure: 18-60 months in the current role (long enough to have shipped, not so long that a move is unlikely).

Only must-haves that map to current-role fields (title, seniority, location, company size, industry) go into the search. Everything else becomes a filter or a yes/no question later. Ask Claude Code to turn those fields into one search and the rest into yes/no questions.

Step 3: Build the candidate pool from people data

Ask Claude Code to run a standard search and keep the session open, so you can refine it in later turns. For this role, ask for senior software engineers at New York fintechs with 50-500 employees.

The reply includes a dataset, a funnel of counts and the filters the agent chose (title variants, industries, a per-company cap). Read the filters before you trust the count: "fintech" is often matched on company descriptions rather than a single industry code. Refine in the same session, for example by adding staff and principal engineers. Decision rule: under 100 people, widen title variants or geography; over 2,000, add a company filter. Cost: none for a standard search.

Step 4: Narrow the pool with filters, then yes/no qualification

Start with free, structured filters. A people dataset carries fields such as skills, management level and months in the current role. Ask Claude Code to preview a filter, such as Python in the skills list and 18-60 months in the current role, then save the result as a new dataset.

Then use AI qualification for criteria that need reading, not matching. Questions must be yes/no about the current role, and several questions are combined with AND logic. For this role, two questions do the work: does the current role involve building backend services, and is the person an individual contributor rather than a people manager?

Cost: qualification spends credits per answered row. Claude Code shows a credit quote before it runs, and it should go ahead only after someone approves that quote. Above 100 rows, qualify in batches of 100. If under 20% pass, the questions are too strict or the pool is wrong; above 90%, they are not filtering anything.

Step 5: Map the target companies

In-house teams use this step to pick the companies to source from; agencies also use it to find companies hiring for the role they place. Run a company search with the same industry, location and size filters, which is free.

Company rows carry the number of active job postings, technical headcount and the latest funding round, so Claude Code can rank them by engineering team size. The postings count is company-wide, not per role. Role-specific postings ("posted a senior backend engineer role in the last 60 days") and headcount growth rates need a research-mode search: the first turn drafts the plan and quotes credits for free, the second spends them, so only approve the second turn after review. The hiring signals guide covers that pattern. To source only from the mapped companies, ask Claude Code to limit the people search to the companies in that dataset.

Step 6: Dedupe the shortlist against your ATS

Remove people who are already candidates, in an active process, or opted out, before you pay to enrich anyone. Claude Code handles this in three moves:

1. Upload the ATS export.
2. Onboard the upload so its rows can be matched, and wait for that run to finish.
3. Drop every shortlist row whose LinkedIn URL already appears in the ATS export.

Check which columns your export uses first, since they depend on your ATS. Agencies that keep candidates in HubSpot or Salesforce can also suppress anyone the connected CRM marks as do-not-contact. Cost: none of these steps are on the published list of credit-spending actions.

Step 7: Enrich contact details where appropriate

Enrich only the deduplicated shortlist, and only the channel you will use. If first contact goes through LinkedIn, skip this step. For email, get a credit quote for the row count first, then run email enrichment only.

Each row needs a LinkedIn URL, or a first and last name plus a company. The output adds an email address and a per-row enrichment status; rows still pending after the one-hour provider window are not charged. Cost: email and phone enrichment are separate per-contact credit lines, so asking for email alone avoids paying for phones you will not call. See the contact enrichment waterfall guide for how providers are stacked.

Step 8: Hand off to outreach and set the cadence

Ask Claude Code to publish the final list and download it as a CSV. Import it into your ATS or sourcing CRM with a source tag and the req id, so replies and hires trace back to the run. Ask Claude Code to draft a first message per row that cites one specific fact from the profile, saved as drafts for a recruiter to edit and send. Keep a short log in the req folder of each search and dataset used, and rerun the search when the req changes or every two weeks while it is open.

Worked example: senior engineers at New York fintechs

The counts below come from one standard search run on November 3, 2026, plus free previews of the same dataset. The last three bullets are an illustrative example.

• Search: a search for senior software engineers at New York fintechs with 50-500 employees returned 1,647 matching contacts across 254 companies headquartered in New York State, with 630 rows saved at a cap of three people per company.
• Free filters: of the 630 rows, 242 listed Python in their skills, 289 had 18-60 months in the current role, and 108 met both conditions.
• Qualification (illustrative example): the two yes/no questions from Step 4 pass 60 of the 108.
• ATS dedupe (illustrative example): 12 are already in the ATS, leaving 48 to verify by hand.
• Enrichment and handoff (illustrative example): after the reviewer drops profiles that no longer match, email enrichment runs on the remaining 40 and the recruiter sends personalized messages to those with a verified address.

The free steps cut the pool by 83% before a single credit was spent. That ordering is the main cost control in this workflow.

How to measure whether it worked

• Profile accuracy: the share of a 20-profile sample that truly matches the brief on review. Hold outreach until it reaches 80%.
• Pool-to-shortlist rate: people sent to outreach divided by the original pool. Track it per req to see which filters do the work.
• Reply rate and interested rate: per req and per message variant, measured in your sequencer or ATS.
• Submittal and interview rate: sourced candidates who reach the hiring manager, and who reach an interview.
• Credits per shortlisted candidate: qualification and enrichment credits divided by people contacted.

Set a baseline on your first three reqs, then review these numbers monthly with the hiring managers you support.

The honest limits of Claude Code for recruiting

• Past-employer searches are weaker than current-role filters. Queries such as "engineers who used to work at a payments company" return fewer and less precise matches, because the strongest filters key on current title and current company. Treat work history as a qualification question and verify it on the profile.
• Profiles go stale. People change roles and the data lags behind. The job change frequency statistics show how often that happens. Check each profile before outreach.
• Skills fields are self-reported. One row in the example listed React and Node.js alongside marketing strategy and Microsoft Word. Use skills as a filter, not as proof.
• Privacy and consent rules apply. Under GDPR Article 14, when you collect data about an EU candidate from another source, you must tell them at the latest at your first communication, and within one month at most. Document your lawful basis and honor opt-outs. In California, the CCPA's exemptions for employment-related and B2B personal information expired on December 31, 2022. Check with counsel for your jurisdictions.
• Do not filter on protected characteristics. Some datasets include an estimated age column. The Age Discrimination in Employment Act protects applicants 40 or older in hiring, so leave such columns out of every filter and qualification question.

Common Claude Code recruiting mistakes

Putting every requirement into the search

Recruiters copy the whole job description into one query because it feels thorough. The pool shrinks to a handful and misses strong candidates whose titles are worded differently. Search on three or four current-role filters, then narrow with filters and qualification.

Enriching before deduping

It is tempting to get emails for everyone first. You then pay to enrich people already in your ATS or in an active process. Run Step 6 before Step 7 every time.

Trusting the count without reading the filters

A big number looks like a good result. If the agent matched "fintech" too broadly or dropped a title variant, the count is wrong in a way you only see in the filters it reports. Read them, and preview 20 rows before moving on.

Sending templated messages at volume

Automation makes it easy to send 500 identical messages. Candidates ignore them and the agency's name suffers. Keep drafts personal to one fact on the profile, and keep a human on send.

How to do this in Landbase

Landbase is the data layer in Steps 3 to 7: describe the candidate or company profile in plain language and get a dataset you can filter, qualify, dedupe and enrich, drawn from 800M+ contacts and 40M+ companies with 1,500+ fields. Run it from Claude Code or Codex through the Landbase CLI or MCP. The same pattern runs marketing workflows and RevOps automation.

Frequently Asked Questions

Can Claude Code source candidates on its own?

Not well on its own. Claude Code can plan the search, write filters and draft messages, but it needs a people-data source to return complete candidate lists. With web search alone it finds scattered examples. Connected to a data layer through a CLI or MCP server, it returns every profile that meets the filters, which a recruiter then reviews.

Which steps of a Claude Code recruiting workflow cost credits?

Standard searches, dataset previews, filters, uploads and ATS deduplication are not on the published list of credit-spending actions. AI qualification, contact enrichment (email and phone are billed separately) and research-mode searches spend credits. Each is quoted before it runs, so approve the quote before confirming.

Is AI candidate sourcing legal under GDPR?

It can be, if you follow the rules. You need a lawful basis, usually legitimate interests documented in an assessment, and you must tell EU candidates where you got their data at the latest at first contact. Keep only what you need, honor opt-outs and delete stale records. Get advice from counsel for your situation.

Does this work for recruiting agencies?

Yes, with two additions. Agencies use the company search in Step 5 for business development, finding companies with open roles in their niche. They also need an exclusion list of client companies, applied before outreach, so no one approaches a client's employees. Run each client req in its own project folder.