How to Use Claude for Talent Mapping (2026 Guide)

A talent map built with Claude lists every person in your target roles, organised by company, level, location and tenure.You start from the companies that employ that talent, pull the people in the target roles at each one, organise them into a grid a client can read, and leave a watch running so the map takes in movers on its own.

Published

Written by

Chris Pisarski

Reviewed by

Manmohit Grewal

Read time

18

minutes

Where the data decides and where Claude reads in a talent map: the data finds companies in the band, people with the title, counts for the grid and someone who moved; Claude reads which companies are comparable, who sits at the real level, the reason on each row and whether a new match belongs.

A talent map built with Claude lists every person in your target roles, organised by company, level, location and tenure.You start from the companies that employ that talent, pull the people in the target roles at each one, organise them into a grid a client can read, and leave a watch running so the map takes in movers on its own.

Now, the split that makes this work is simple: the data does the counting, and Claude does the reading. A filter can tell you which companies sit in your client's industry and headcount band, and it can tell you who carries a Director title. What it can't tell you is which of those companies your client loses engineers to, or whether a Director at a 40-person startup is the level your client actually means. That's where Claude comes in, reading the speciality tags, the company's own site and each work history, deciding what belongs on the map, and writing the reason next to every row.

So the split is clean. Claude judges comparability and level, and writes the reasoning. The people and company data supplies the companies, the titles, the counts and the alert when somebody moves and you hold the brief.

To see how far the filters get on their own, I ran the company step on one market: US software companies with 200 to 2,000 staff, carrying a speciality tag containing "payments". 94 companies matched, and I read ten of them. Every payments tag was accurate, and yet their other tags named dental and optometry practices, online reviews and blockchain nodes. Sorting those ten into "comparable" and "not really" is Claude's job on this page, and it does it by reading each company's own words.

Quick answer: how do you use Claude for talent mapping?

Write a role brief that separates must-haves from nice-to-haves, use Company Search to find the companies that employ that talent, and have Claude keep only the ones that are genuinely comparable to your client. Then pull the target roles at those companies with People Search, have Claude place each person at their real level with a one-line reason, and lay the result out as a grid by company, level, location and tenure. Two watches keep the map up to date as people move.

At a glance

  • Where it starts: the companies that employ the talent, before any person is pulled.

  • Claude's job: judging which companies are comparable, reading work histories for real level, writing one line of reasoning per row.

  • The data's job: the company set, the titles, the counts, and the alert when a mapped person moves.

  • What you need: a role brief, the client's domain, and Claude with the connector.

  • What a first build costs: about 171 credits for 40 companies and 600 people, at the published list rates.

  • How it stays up to date: a discovery watch on the map's filter, and an entity watch on the people already mapped.

  • Best fit: a search firm building a map a client will see, or an in-house team mapping a competitor's function.

  • Main limit: a vague brief produces a wide map nobody can act on.

What is a talent map, and how is it different from a candidate list?

A talent map counts and groups the people who match a role, while a candidate list names the ones you're going to call.

The map answers how many qualified people exist, and which companies they cluster in, and how quickly that population thins out as the brief gets stricter. That work happens before sourcing, and it's what lets you tell a hiring manager their spec leaves twelve people in the metro they want.

Talent mapping and market mapping are the same job under two names. Both mean the population behind a role.

Can't I just ask Claude for the map?

No, because the model can't be the data layer, and two recruiters have told us exactly what that feels like in practice. A recruiter at an executive search firm who has used Claude and ChatGPT directly for this summed it up as "You may have a few names, but never thorough enough." A partner at a venture firm who recruits for portfolio companies put it as "Claude does okay, but not right" for talent mapping and competitor research.

The reason is straightforward. The model answers from what it learned at training time, while a map is a claim about who works where today. Claude's training data stops at a cutoff, which Anthropic documents for each model, so names and current employers have to come from a database you query as you build the map.

A filter-only map fails the other way round, which is what step 2 walks through.

How do recruiting teams already do this with Claude?

Market and talent maps came up on our calls with 46 organisations, mostly agency and executive search firms building a map a client will see, alongside in-house talent teams mapping a competitor's function. Three patterns stand out, and they differ mostly in how much of the map is standing work.

The client document, assembled in a chat. An owner at a recruiting firm built a market map during a client search using the Claude connector to Crustdata, and then had Claude write the one-page document he shared with the client. His review: "it really made everything look nice, but it was the data behind it that made it effective."

The two-week map, by hand. A recruiter at an executive search firm described the manual version to us. He finds companies of similar size and scale doing exactly what the client does, breaks those down, and finds who the CFO is and who might be next in line. He puts it at two weeks minimum per map, which is time he'd rather spend on conversations.

The standing map that takes in movers. A founder who runs searches for high-growth startups keeps a standing map, and he told us he doesn't want to spend his time watching how the universe changes. He wants the people who newly fit to arrive in the map on their own.

In-house teams run the same play on a competitor. A talent lead at an AI infrastructure startup described counting a large chipmaker's compiler engineers and then working out which handful were worth approaching first.

The guide below follows the first pattern and then adds the third: you build the map in a chat with the connector, and a watch keeps it up to date without you.

What does this guide cover, and what does the hand-built version cover?

This guide is the Claude build, and the hand-built guide is the manual method. Neither repeats the other.

The job

Where it's covered

Reading a job description down to filters

The hand-built guide

Sizing a population with free counts

The hand-built guide

Breaking a total down by company, title and geography

The hand-built guide

Judging which companies belong on the map

This guide

Telling a real Director from a Director in title only

This guide

Writing the reason that makes the map client-ready

This guide

Keeping the map up to date as people move

This guide

Connecting Claude to the data

Recruiting workflows with Claude

Ranking and scoring the people on the map

Multi-layer candidate filtering

The map itself, from the companies through to the refresh, is what this guide covers, and it takes about fifteen minutes to read. The hand-built version is in our guide to building a recruiting market map from a job description.

What do you need before you start?

Four things, and only one of them is software.

  • A brief with the must-haves separated out. The filters are built from the must-haves, and a wish list dressed up as requirements produces a map of nobody.

  • Your client's domain, and any competitors they've already named. Company Identify resolves the domain to a company record, which gives you the industry and size band to match against.

  • Claude with the Crustdata connector, or an API key if you're writing the code yourself. A new self-serve Crustdata account comes with free credits.

  • A number you're willing to spend. Counts are cheap and rows are not, so set the budget first.

How do you connect Claude to the data?

Three routes reach the same search tools, and the connector on claude.ai is the one that needs no code.

Which route fits a map you build once in a chat?

Use the connector on claude.ai. Add the hosted Crustdata MCP server as a custom connector, then describe the role and let Claude call the search tools in the conversation. This is the route to start on.

Which route fits a map you want to rebuild on demand?

Use Claude Code with the Crustdata plugin. The Crustdata CLI installs the plugin, which carries the MCP server and the skills, and the map build becomes a prompt you can re-run.

Which route fits a map inside your own product?

Call the API from your own code and give Claude the results as tool output. You keep the brief, the grid and the refresh in a repository, which makes sense when the map is a standing asset. The sourcing side of that build is covered in recruiting workflows with Claude.

How does a talent map come together in Claude?

The map comes together in five passes, and each one hands the next something you can check:

  • Your brief becomes a company profile and a people profile.

  • A company search returns the companies that fit the profile.

  • Claude reads those companies and keeps the comparable ones.

  • A people search returns the target roles at the companies that survived.

  • Claude places each person at a real level and writes the reason.

Steps 1 to 5 below follow these passes, and step 6 adds the refresh. The division of labour matters more than the order, because it keeps Claude on the work a filter can't do. Every step a filter settles alone sits with the data, and every step that needs somebody to read something and form a view sits with Claude.

Where the data decides and where Claude reads in a talent map: the data finds companies in the band, people with the title, counts for the grid and someone who moved; Claude reads which companies are comparable, who sits at the real level, the reason on each row and whether a new match belongs.

The pass

Who does it

What it produces

Brief to profile

Claude

A company profile and a target-role profile

Company set

Company Search

Companies in the industry, size band and geography

Comparability

Claude

A kept or dropped verdict per company, with a reason

The people

People Search

The target roles at each kept company, with work history

Level and fit

Claude

A band per person and one line of reasoning

The grid

Your code or the chat

Counts by company, level, location and tenure

Refresh

Watchers, then Claude

New matches checked against the brief before they join

Step 1: How do you write the role brief the map is built from?

Write the brief as three lists, because the map inherits whatever vagueness you leave in it.

Let's say your client is a 400-person US payments company and therole is Director of Data Engineering, owning the ledger and reporting platform, reporting to the VP of Engineering. The brief goes to Claude like this.

  • Must have run a data platform team of six or more. An analytics team doesn't count, and neither does a staff engineer with no reports.

  • Must have owned batch and streaming in production at a company that moves money. Payments, lending, banking and exchange all count.

  • Must have been the data owner through a compliance audit. PCI or SOC 2 both count, and their name has to be on the evidence.

  • Nice to have: took a platform through a tenfold volume increase. This line separates the top of the band.

  • Nice to have: hands-on with the client's stack. This one is useful in a conversation and it is never a filter.

  • Disqualifier: analytics or BI leadership with no platform ownership. The titles look identical from outside the company.

  • Disqualifier: consulting or agency only. They have advised on platforms without carrying one.

  • Geography: the New York metro, plus remote US for the right person. Anywhere else is a conversation with the client before it is a filter.

Three of those lines go straight to filters. They are the seniority band, the geography and the employer type. Everything else needs a reader.

Step 2: How do you find the companies that actually employ this talent?

Start from the companies your client competes with for people, in two passes.

The first pass is the competitor field. Company Search carries competitors.websites as a filter, so you can ask for companies whose competitor set already includes a given domain. I ran it against one public payments domain, inside the US and the 200 to 2,000 band. It returned four companies, Plaid, Braintree and Paystack among them. That's a useful start, but it's narrow, so treat it as a first pass and widen it.

The second pass widens it. Match the industry, the headcount band and the geography from your client's own record, then add a speciality tag for the market:

{

"filters": {

"op": "and",

"conditions": [

{"field": "taxonomy.professional_network_industry", "type": "=", "value": "Software Development"},

{"field": "locations.country", "type": "in", "value": ["USA"]},

{"field": "headcount.total", "type": "=>", "value": 200},

{"field": "headcount.total", "type": "=<", "value": 2000},

{"field": "taxonomy.professional_network_specialities", "type": "contains", "value": "payments"}

]

},

"fields": ["crustdata_company_id", "basic_info.name", "basic_info.primary_domain", "headcount.total", "taxonomy.professional_network_specialities"],

"limit": 10

}

The filter above matched 94 companies when I ran it. Ten came back as Bolt, Recharge, Airbase, Podium, Weave, EasyEcom, Blockdaemon, Jumio, Incode and HoneyBook. Every one carries a payments speciality tag, and every tag is accurate. Its the rest of their tags are what separate them. Podium's other tags name online reviews and reputation management, Weave's name dental and optometry practices, and HoneyBook's name booking for creative businesses.

So the tag tells you a company touches payments. It doesn't tell you whether your client loses engineers to it.

How does Claude decide which of the 94 companies are comparable?

Claude keeps a company only where its own account of itself says money movement is the product. Give it the text to work from, because the tags alone won't carry that judgement. Company Search returns the speciality and category tags. For anything thinner, Web Fetch reads the company's own site at 1 credit a page.

  • Give Claude the whole brief. A company is comparable when it hires the same engineers for the same problem. The judgement needs the must-haves in front of it.

  • Ask for a reason per company, kept or dropped. That's what makes the map defensible when a client asks why their biggest rival is missing.

  • Keep the dropped list. A company you dropped for the ledger role may be the right comparable next time.

Step 3: How do you pull the people in the target roles at each company?

Filter People Search on the company ids from step 2, never on the company names.

The company ids matter, and skipping them cost me a correction mid-run. I filtered on six employer names, with a title matching "data" and a seniority band of Director, Vice President or CXOand got 11 people back. The same filter with the crustdata_company_id values in place of the names returned 7. Four of the eleven carried the employer name Bolt, threeof those sit in Tallinn, with titles naming ride-hailing and food-delivery work, and the US checkout company on my list is a different company altogether.

Company names collide and company ids don't, so carry the id through from the company search:

{

"filters": {

"op": "and",

"conditions": [

{"field": "experience.employment_details.current.company_id", "type": "in", "value": [631273, 631818, 632011, 633428, 635135, 711163]},

{"field": "experience.employment_details.current.title", "type": "(.)", "value": "data"},

{"field": "experience.employment_details.current.seniority_level", "type": "in", "value": ["Director", "Vice President", "CXO"]}

]

},

"fields": ["basic_profile.name", "basic_profile.location.city", "experience.employment_details.current.name", "experience.employment_details.current.title", "experience.employment_details.current.seniority_level", "experience.employment_details.current.start_date"],

"limit": 100

}

Two details in that query are worth knowing before you run your own.

  • seniority_level takes exact values. They are Owner / Partner, CXO, Vice President, Director, Experienced Manager, Entry Level Manager, Strategic, Senior, Entry Level and In Training. Send one of those strings, and write Vice President rather than VP.

  • The (.) operator matches all the words in any order. Use it for title variants. The same job is called data engineering at one company and data platform at the next.

Widen the title before you narrow the level. Filtering those six company ids for "data" anywhere in the current title returned 100 people and seven of them sat in the Director, Vice President or CXO bands. That funnel from 100 to 7 is the map taking shape, and both numbers are worth showing a client.

Step 4: How do you tell a real Director from a Director in title only?

Hand Claude the work history and ask it to place the person, because the title alone won't do it.

Of the 7 people that filter returned, two titles described owning a data platform: Head ofHead of Data Platform and Director of Data Engineering. The rest named analytics, data product management and AI leadership. Against a brief whose first must-have is platform ownership, that's two rows to work and five to park, and no filter in the set could have told them apart.

The fields behind that judgement sit in the search response. Start dates give you tenure in the role and at the company, and past roles with their dates give you the shape of the career. Job Title Normalization , currently in beta, maps each raw title to a department, a category and a standard matched title, so you can compare the same job across companies that name it differently. It reads functional intent and not seniority, so the level judgement stays with Claude.

  • Ask for a band and a reason. A reason such as platform owner, ran a six-person team through a payments migration, is reviewable. An 8 out of 10 is not.

  • Tell Claude which judgements aren't binary. A founder who runs these searches calls a boomerang move a weak signal, and not a rule. Somebody may have gone back when their new employer fell apart. An agent given that as a rule applies it to everybody.

  • Read the role that matches the brief. A person can hold more than one current role, and one in my pull held two at the same employer after a promotion. Match the role to the brief before you read its start date.

Step 5: How do you organise the map by company, level, location and tenure?

Lay the map out as one row per person, with four columns a client can sort on: the company, the level you placed them at, where they work, and how long they've been there. Tenure comes from the role start date, so a years-in-role column falls out of the data you already pulled. Counts by company and by level come from the same filter with the result limit set to one. That's the cheapest call on the page.

A client-ready map is small. The owner at the recruiting firm who used the connector had Claude write his as a one-page document.

  • Lead with the funnel. Show the population at the broad title, then at the real level, then in the geography the client wants.

  • One line of reasoning per row. Claude writes it in step 4, and it is the column a client actually reads.

  • Keep the counts beside the names. The counts are what a client reads first, and they carry the funnel.

The breakdown mechanics, including the free sizing counts and the geography splits, are covered in the hand-built market map guide.

Step 6: How do you keep the map up to date as people move?

Run two watches, because a map goes out of date in two directions: people who newly match your brief have to arrive. People already on the map move on. A discovery watch handles the first and an entity watch handles the second:

Two watches keep the talent map up to date: a discovery watch finds people who newly match (0.5 credits a new person) and an entity watch follows mapped people who move (5 credits a changed person); Claude checks each against the brief before it joins the map by company, level, location and tenure.

Both watches report to Claude before anything reaches the map. Claude checks each new match against the brief that built the map, and each move by a mapped person against the same brief.

  • A discovery watch re-runs the map's own filter. It takes the same filter syntax as the search endpoint, and delivers the people who newly match. The first run is a free baseline capped at five records. Later runs cost 0.5 credits a new person at the default 30-day refresh.

  • A person entity watch tracks the people on the map. You give it a fixed list of up to 10,000 people. You add a track tree for the change you care about which is for a map that change is the employer and the title. It fires once per change, at 5 credits a changed person on the 30-day refresh.

  • Claude reads every new match before it joins the map. A filter that suited the market in March pulls in people the brief would reject. That check is what keeps a refreshed map worth reading.

The two watches answer what the founder running standing searches asked us for. In his words: "I want to understand that as that universe changes, those people are being ingested into my market map."

Keep the filter tight. A discovery watch whose filter matches more than 500,000 records is refused at creation, which is the system telling you the map is too wide. Routing alerts into Slack, a CRM or an ATS is covered in our guide to tracking job changes.

How much does a talent map cost to build and keep fresh?

A first build of 40 companies and 600 people costs about 171 credits at the published list rates, and keeping it fresh costs roughly 160 credits a month on top of that.

Search endpoints charge per result, and a premium field group bills once when you filter on it and again when you receive it. That's why your choice of columns matters more than your row count.

The call

Rate

This map

Credits

Company Identify

Free

1 client domain

0

Company Search with headcount and taxonomy

0.63 per result

40 companies

25.2

People Search with work history

0.23 per row

600 people

138

People Search, counts only

0.13 per query

60 grid cells

7.8

Discovery watch, 30-day refresh

0.5 per new person

20 a month

10

Entity watch, 30-day refresh

5 per changed person

30 a month

150

Crustdata credits run from $0.30 to $0.01 per credit, lower depending on your plan, so the same map sits in a very different place on a self-serve account and an enterprise one. Your account's own rates are in the pricing docs.

How can you cut the bill?

Count before you pull, and pull fewer columns.

  • Size every cell with the result limit at one. A count costs 0.13 credits against 0.23 for the row.

  • Drop the field groups you aren't reading. A company row with no headcount and no taxonomy costs 0.03.

  • Stay on the 30-day refresh unless the map feeds outreach. The same new person costs 0.5 credits at 30 days and 150 at one day.

  • Let the watch do the re-running. A discovery watch bills for the people who newly match. A monthly re-run of the whole search bills for everyone who still matches. Where do I start this week?

Pick one open role and build the map for it.

  • Write the brief as three lists. Separate the must-haves, the nice-to-haves and the disqualifiers, and mark the lines a filter can settle.

  • Resolve the client's domain and run the two company passes. Start with the competitor filter, then widen with the industry, size and speciality filter.

  • Have Claude keep or drop each company with a reason. Keep the dropped list for the next search.

  • Pull the target roles at the survivors by company id, then count the grid. Run the counts first, and pull rows only where you need names.

Start free at crustdata.com, where a new self-serve account comes with free credits, or see the workflows the map feeds in our recruiting solutions.

Talent mapping FAQ

What is talent mapping?

Talent mapping counts and groups the people who could fill a role, by employer, level, location and tenure. It happens before anyone is contacted, and it tells you whether a role is fillable and where the talent sits, which is a different question from who to call first.

How long does a talent map take to build?

A manual map takes two weeks at least at an executive search firm. That's the account one recruiter gave our team. The Claude version front-loads the brief and the comparability read, and the searches run as you talk to it.

Can Claude build a market map without a data provider?

No, because a map is a claim about who works where now, and the model answers from training data. Claude needs a people and company search it can call as it builds the map and the judgement it adds on top is what makes the map worth reading.

What's the difference between a market map and a talent map?

There isn't one. Market map and talent map are the same exercise under two names, and neither means a shortlist.

How do you keep a talent map up to date?

Run a discovery watch on the map's own filter for people who newly match. Run a person entity watch on the people already mapped, for the ones who change job or title. Both bill only for what they deliver, so a quiet month costs nothing.

Does a talent map include salary data?

A talent map built this way carries no salary data. It covers employer, level, location and tenure, and compensation benchmarking is a separate source you bring yourself.

Crustdata is now live in Claude!

Find the best leads, candidates and companies with only a prompt.

Sign In