How to Set Up Job Change and New Hire Alerts in Claude (2026 Guide)
You need two watches to cover job change and new hire alerts, one on the people you already know and one on the accounts you care about.
Published
Written by
Chris Pisarski
Reviewed by
Manmohit Grewal
Read time
20
minutes

You need two watches to cover job change and new hire alerts, one on the people you already know and one on the accounts you care about.
The first fires when someone on your list starts a new job, and second fires when anyone newly matching your filter shows up at one of your accounts. Both can deliver to a webhook, a Slack channel or an inbox, and both charge you only when something actually fires.
Now, here's the split the rest of this guide is built on: the data does the detecting, and Claude does the reading. A watch can tell you that a contact moved, which employer they moved to and when they started, but four questions stay outside its reach:
Does this particular move matter to this account?
Have we spoken to this person before, and how did it end?
Does the person still own a budget you sell into?
Which four of this morning's fifty alerts should a person open first?
A watch can't write the message either, which makes five jobs in total, and all five are Claude's work on this page.
Quick answer: how do you set up job change alerts in Claude?
Create a person entity watch on the contacts you already track and a person discovery watch on your target accounts, then send every alert to Claude along with a short relevance brief that says what you'd act on. Claude returns a tier, a one-line reason and, for the alerts that earn one, a draft message that a person reviews and sends. Start on the 30-day freshness tier, preview every watch before you create it, and key the track so promotions stay quiet.
At a glance
What fires on a known list: a person entity watch, at 5 credits per changed person on the 30-day tier.
What finds people you don't know: a person discovery watch, at 0.5 credits per new person.
What the first run costs: nothing. It records a baseline, fires on nothing and sends a free sample of up to five.
What a quiet run costs: nothing. You pay per delivered notification.
Claude's job: judging whether a move matters to the account, ranking the morning, drafting the message.
The data's job: detecting the change, the employer, the title and the start date.
The main limit: a watch reports what the last refresh saw, so freshness is the knob that sets detection speed. Which knob controls how fast you hear about a move?
The one thing worth settling before you build anything is which knob controls speed, because it isn't the one most people reach for first. You pay per notification, and the freshness tier you pick sets both the price and how quickly a real move gets detected. Checking more often does nothing on its own, since a watch can only report what the last refresh saw.
For what it's worth, testing all of this cost me nothing. I previewed six watches and created one for real while writing this guide, and the whole run came to zero credits. A preview of a watch for US executive leadership who recently changed jobs came back with 47,660 people before anything was created, and the watch I did create resolved all three people on its list, delivered them as a free sample and charged zero credits.
What are the two ways to build this?
The connector in a chat. You describe the watch, Claude creates it, and the alerts come back to the same conversation. This is where most people should start.
Your own code against the API. You keep the relevance brief, the routing and the drafts in a repository, which makes sense when alerts feed a product or a whole team.
What you'll need:
Claude with the Crustdata connector, or an API key if you're writing the code yourself
A Crustdata account, which comes with free credits for new self-serve users
A list of people worth watching, and a list of accounts worth watching
A written rule for what you'd actually act on, which step 1 builds
This guide covers the judgement that happens after an alert fires, and it takes about fifteen minutes to read. If you want the hand-built version, with the webhook handler and the routing code, that lives in our guide to tracking job changes.What does this guide cover, and what does the hand-built version cover?
Everything after an alert lands is covered here, and everything before it is covered in the hand-built guide.
Both use the same two endpoints, and they part company at the moment the notification arrives, because that is where a filter stops helping. No field says whether a VP of Platform at a 400-person payments company is worth waking a rep for.
The question | Covered here | Covered in the hand-built guide |
|---|---|---|
Creating the watch by describing it to Claude | Yes | No |
Writing the create call by hand | Yes | Yes |
Webhook endpoint and signature validation | No | Yes |
Parsing the delta payload in code | No | Yes |
Routing rules to Slack, a CRM or an ATS | No | Yes |
Deciding whether one alert matters to one account | Yes | No |
Ranking a morning of alerts | Yes | No |
Drafting the message that goes out | Yes | No |
How do sales, recruiting and investing teams already do this with Claude?
Four patterns came up on our calls, and each one puts a different job between the alert and the person reading it.
The team that built it beside the tools they already pay for. A GTM leader at an AI software company told our team about their go-to-market engineering team where a few people there had put something together with Claude covering champion tracking and job changes, alongside the licences they already held.
The team that wants a draft attached. An engineering leader at a B2B software company already runs Claude on a virtual machine every day and diffs his own queries. What he wants from a watch is for Claude to take the result into the CRM and write the email. In his words, "we don't have to even draft the email, we just have to hit send".
The recruiter who reads the direction of the move. Someone doing recruiting business development watched about 6,000 US startups for anyone joining in the software engineering discipline, which returned a few hundred people a month. What he acted on was the direction of the move: a joiner who arrived from another employer meant there was now an opening at the employer they had left. boards.
The recruiting firm that watches the client companies. Someone building recruiting automation told our team that watching thousands of candidates was the wrong way round, since many candidates map to few clients. His version: "whenever they get a new hire, we can just check if that hire is somebody we've already submitted to them".
This guide follows the second pattern, because it's the one that ends with a person reviewing a draft.
What does the investing version look like?
An investor watches departures from one company, where the sales reader watches arrivals at many.
An investor at a venture fund toldus how his firm catches it today: theydiff yesterday's check against today's, looking for someone leaving a frontier AI lab for a stealth company. The second thing he wanted was narrower. In his words, "if this person is someone we're close with and they just moved out of a portfolio company into another company, we just want to have a heads up."
Both are the same two watches, pointed at a list the fund already knows and a filter describing the lab. Another fund we spoke with set one up with Claude Code, which is the second route in the section on connecting Claude below.
How does the alert loop work?
The loop has four stages, and each hands the next something it can check:
A watch detects a change and delivers it.
Claude reads the change against your written brief.
The alert gets a tier, a reason and, where it earns one, a draft.
A person reviews what reached them and sends.
Each stage settles a different question. Only the middle band needs the brief you'll write in step 1, since the band above it is a field lookup and the band below it is routing.

The step | Who does it | What it produces |
|---|---|---|
Detecting the change | The people data | A change event with the old and new values |
The surrounding profile | The people data | The field groups you requested |
Suppressing a same-employer promotion | The watch config | Silence, from the keys list on the track |
Judging whether they left the old job | Claude | A yes or no, read off the employment history |
Checking the territory and the headcount band | The company data | Industry, country and size for the new employer |
Reading what happened with this person before | Claude | The deal history, as context for the tier |
Judging relevance to the account | Claude | A tier, and one line saying why |
The message | Claude | A draft a person can edit and send |
Delivery | Your webhook, Slack or email | The alert, where the team already works |
The decision to contact | A person | The sent message |
Can't I just read the alerts myself?
You can, right up until the volume makes it a chore nobody does and the teams we talk to describe the same problem.
Someone building job-change tracking at a GTM data company put the problem plainly : awatch on a known contact returns that contact's move, "but your job change might be completely irrelevant to us. Because maybe your job change was to a senior software engineer, but we are looking for chief executive officers."
Someone building contact tracking at a sales-intelligence company described the same thing as a hierarchy, where customers"care about people within the personas that they sell to the most", then executives, and the rest not at all.
Which watch fires on people you already track?
An entity watch fires when a tracked field changes on a list of people you supply.
You give the person entity watch up to 10,000 people and a track condition and for a job change that condition is experience.employment_details.current with the type added. The list is editable at any time. Anyone you add later is baselined quietly on the next run, so adding someone never sets off a false alert.
Which watch finds people you don't know yet?
A person discovery watch re-runs a saved search and delivers whoever newly matches.
This is the one you want for new hires at your accounts, because a new hire is, by definition, nobody you were tracking. The filter grammar is the same as Person Search, and one condition does most of the work here: recently_changed_jobs, a boolean that marks a past role which ended within 90 days of the last index rebuild. To find people who recently started a role, filter experience.employment_details.current.start_date on a date range.
How fast will you hear about a move?
A move surfaces within the freshness tier you set, which runs from 30 days down to one.
The cadence has nothing to do with it, and the docs put this plainly:"Raising every_hours does not make a change surface sooner on its own. The watch can only report what the last refresh saw." Freshness controls how fast a change is detected, and the tier you pick is what you pay for. The cadence only controls when you hear about it.
refresh_frequency_days | Credits per changed person |
|---|---|
Not set, or 30 | 5 |
14 | 10 |
7 | 20 |
3 | 50 |
1 | 150 |
Most teams should start on the 30-day tier and move up one step at a time. A champion who moved last week is still worth a message this week.
Why does a promotion look like a move?
Because a new title at the same employer arrives as a new element in the same array, exactly the way a move does.
A bare added on experience.employment_details.current fires on a promotion as readily as on a move, and you pay for both. A second role alongside the first does the same thing. A technical buyer at a sales-intelligence startup had already written his own check for this, telling our team his system asks "is this a real job change? Are they still current in their previous role?"
This isn't an edge case either, since two of the three people on the watch I created carry more than one current role. The fix is a keys list on the track leaf, which names the fields that identify an element:
{
"field": "experience.employment_details.current",
"type": "added",
"keys": ["crustdata_company_id", "name"]
}
Now the watch fires only on an element whose company id and name are both new to it. The docs add two cautions:
The loosest entry wins. A bare name entry would defeat a composite one, so give a fallback entry the same fields as the entry it backs up.
A return to a former employer stays silent. The watch remembers every identity it has seen, and this one is deliberate.
keys leaves one case for Claude. It silences a new title at an employer the watch has already seen. It says nothing about someone who takes a role at a new company while keeping the old one. Only the employment history in the payload says whether the previous role has ended, which is the question the buyer above was asking.
How do you connect Claude to Crustdata?
Three routes reach the same watch endpoints, and the first one needs no code at all.
Which route fits a rep who doesn't write code?
Add the hosted Crustdata MCP server as a custom connector on claude.ai, then describe the watch you want. Claude creates it and reads its runs back to you in the same conversation.
Which route fits a watch you want to rebuild on demand?
Use Claude Code with the Crustdata plugin. Install it with claude plugin marketplace add crustdata/skills and claude plugin install crustdata@crustdata-plugin, then sign in with /crustdata:login. This is the route the venture fund above used.
Which route fits alerts inside your own product?
Call the API from your own code and hand Claude the payload as tool output. The sourcing side of that build is covered in recruiting workflows with Claude, and the rubric pattern for scoring is in multi-layer candidate filtering.
The setup, in six steps
Here's the whole build before we go into each step, so you know where it's heading:
Write the relevance brief, the document Claude reads on every alert.
Create a person entity watch on the people you already know.
Create a person discovery watch on your target accounts.
Choose how alerts reach Claude: pull-only, your own webhook, or a channel the team already reads.
Have Claude tier each alert against the brief.
Have Claude draft messages for the top tier and roll everything else into one morning post.
Step 1 matters most, because every step after it reads the brief.
Step 1: What should the relevance brief say?
The brief says what you'd act on, at what speed, and what you'd never want to see.
It's the document Claude reads on every alert, and it's the one thing no filter can hold, so write it the way you'd brief a new rep. The worked example below is the brief for an account executive at an observability vendor, carried through the rest of this guide.
Territory: US and Canadian fintech, 200 to 5,000 staff.
Who I sell to: the Director, VP or CXO who owns the platform
or observability budget.
Watching 94 people: 21 closed-won champions, 33 contacts on
live deals, 40 from deals we lost or stalled.
Watching 60 named accounts for anyone newly hired into that seat.
Act today:
a closed-won champion lands anywhere in the territory,
in a seat that owns the budget.
Act this week:
a closed-won champion lands in the territory below that seat,
or a contact from a lost deal lands in it.
Log only:
a promotion inside the same employer,
a second role alongside the first,
any move to a company under 200 staff.
Never route to me:
a move out of engineering into a function with no budget,
a move to a company that is already our customer.
That one goes to the CSM.
Four things make this brief usable:
It names the territory in numbers. Country, industry and headcount band are filter fields, so "in territory" is settled before Claude reads anything.
It separates speed from relevance. A champion landing two levels below the budget seat still matters, and it can wait until Thursday.
It says what to suppress. Promotions, second roles and moves to companies under 200 staff never reach the rep.
It routes one case away from the reader entirely. The CSM gets the customer moves, and they stay off the rep's action list.
Step 2: How do you watch the people you already know?
Create a person entity watch on the list, and track the current employment array.
The example below watches contacts by their work email address, tracks a new current role, keys the track so a promotion stays quiet, and posts to a Slack channel. Every request to the watch endpoints needs the x-api-version: 2025-11-01 header.
curl --request POST \
--url https://api.crustdata.com/watch/person \
--header 'authorization: Bearer YOUR_API_KEY' \
--header 'content-type: application/json' \
--header 'x-api-version: 2025-11-01' \
--data '{
"entities": {
"business_emails": ["first.last@example.com", "another@example.com"]
},
"track": {
"field": "experience.employment_details.current",
"type": "added",
"keys": ["crustdata_company_id", "name"]
},
"fields": ["basic_profile", "experience"],
"config": {
"trigger": { "type": "interval", "every_hours": 24 },
"refresh_frequency_days": 30,
"max_results_per_run": 100
},
"notifications": [
{ "type": "slack", "webhook_url": "https://hooks.slack.com/services/T000/B000/XXXX" }
]
}'
Three keys in that body change what you get back:
fields is separate from track. The field you track is not added to the payload automatically. If you omit fields and the delivered record carries only basic_profile and social_handles, so ask for experience if you want the new employer in the alert itself.
track and fields are fixed at creation. You can edit the watched list, the schedule and the channels afterwards, but changing what you watch for means a new watch.
fields does not affect the price. You're charged per notification at the freshness tier you chose, whatever you ask to be delivered.
Send the same body with "preview": true first. The response says whether the watch can be created, how many of your people resolved, and the credits per notification, and it creates nothing. When I previewed a watch on a single work email address, it resolved that person and reported 5 credits per notification, for no charge.
Then read the first run before you trust the channel. The baseline run records a starting value for everyone and fires on nothing, andsends a free sample of up to five people, each with an empty changes array. The watch I created had no channel attached. It resolved all three people, logged "0 credits used" and delivered every entry with changes: []. So areceiver that ignores anything with an empty changes array skips the whole baseline by on its own.
Step 3: How do you watch target accounts for new hires?
Create a discovery watch whose filter describes the seat.
The brief in step 1 wants anyone newly hired into a budget-owning engineering seat. Start from the seat, then narrow to the account list:
curl --request POST \
--url https://api.crustdata.com/watch/person/search \
--header 'authorization: Bearer YOUR_API_KEY' \
--header 'content-type: application/json' \
--header 'x-api-version: 2025-11-01' \
--data '{
"filters": {
"op": "and",
"conditions": [
{ "field": "experience.employment_details.current.title", "type": "(.)", "value": "VP of Engineering" },
{ "field": "professional_network.location.country", "type": "=", "value": "United States" },
{ "field": "recently_changed_jobs", "type": "=", "value": true }
]
},
"fields": ["basic_profile", "experience"],
"config": { "trigger": { "type": "interval", "every_hours": 24 }, "max_results_per_run": 25 },
"notifications": [
{ "type": "slack", "webhook_url": "https://hooks.slack.com/services/T000/B000/XXXX" }
]
}'
The filter above carries no account list yet, and that's deliberate. Once you add the accounts as a condition on the current employer, the reach drops from a whole market to your list.
Preview this one too, because the reach is the bill. A discovery watch whose baseline would follow more than 500,000 people is refused at creation. That refusal is the system telling you the filter is too wide.
Preview also shows you what a loose filter will actually deliver. When I previewed that exact title filter across the US, it reported 72 people and of he three it sampled, one was a VP of mechanical engineering and one an embedded-systems founder. The filter matched the title text it was given, so narrow it with the account list and the seniority band, and send the rest to Claude to read.
Read the count the preview gives you, and treat a zero as a problem to solve before you create anything, because a watch whose filter matches nobody today will just sit there delivering nothing. Filter values have to match the stored vocabulary exactly. The seniority values in the Person Search reference are exact strings, including "Director", "Vice President" and "CXO".
Step 4: How do you get the alerts in front of Claude?
Three delivery shapes work, and the simplest skips the webhook entirely.
Leave the watch pull-only. Send "notifications": [] and nothing is pushed anywhere. You read the results with the run endpoints instead, which Claude can call through the connector. This is how I read the run in step 2, and it needs no receiver at all.
Point the webhook at your own code. Your handler takes the payload and calls the Claude API with it, along with the brief. Pick this when alerts feed a product.
Post to the channel the team already reads. The watch delivers to Slack, Google Chat or email, and whatever agent you keep in that channel picks it up. A founder at a sales software company described exactly that shape to our team. His daily agent reads the channel and hands him one digest, because "no one uses their dashboards".
Whichever route you pick, the brief travels with the alert. Claude has no memory of your territory between conversations. The brief goes in the prompt, a project, or a file your code reads on every call.
Step 5: How does Claude decide who sees an alert?
Claude gets the brief, the alert and a fixed output shape, and returns a tier and a reason.Now the output shape matters more than the prompt wording. Ask for a tier from your own list, one line of reasoning, and the two facts behind them, so your code has something it can route on:
Below is a job change alert and the relevance brief it should be
judged against.
Return JSON with exactly these keys:
tier: one of act_today, act_this_week, log_only, route_to_csm
reason: one sentence, naming the fact that decided it
left_previous_employer: true if the previous role has ended,
false if they are holding both
prior_contact: what the CRM history shows about this person,
or none if there is nothing on file
Rules:
Judge only against the brief. Do not infer a tier it doesn't name.
If you cannot place the new employer against the territory,
say so in the reason and do not guess a tier above log_only.
Never return act_today for an account already marked a customer.
Brief:
Alert:
Two things are worth building in from the start:
Ask for the deciding fact in the reason. A rep can then disagree with the tier without re-reading the payload.
Give Claude a way to say it couldn't place an employer. A confident wrong tier is worse than an honest unknown.
Step 6: How do you get Claude to draft the message and the morning post?
Draft only for the top tier, and collapse everything else into one post.
A sales leader at a large technology company told us he wanted material changes visible "as close to real-time as possible". Two months later, a sales leader at the same company asked for one consolidated post for everything else, because a thread per watch was too much for a channel that was already busy. Both are right, and the brief's tiers are what separate them.
The same morning, seen from both ends:

Two of the fourteen earned a draft, three went into the queue, and the other nine never reached the rep's action list. The post collapses everything that fired since the last one, so a quiet night produces a short post.
For the top tier, hand Claude the alert, the brief and whatever account history you can reach. Ask for a draft a person can edit in a minute, and keep three rules in the prompt:
Name the specific move. The message should be impossible to send to anyone else, which means the old employer, the new seat and what they already know about you.
Leave the ask small. Someone three weeks into a new job is not signing anything.
Stop at the draft. A person reviews and sends, because the brief can be wrong and only they can tell.
How much does it cost to run?
You pay per delivered notification, so the bill tracks movement rather than the size of your list.The worked example watches 94 people and 60 accounts. The table below applies the published list rates to one month of that, and the credit figures are our arithmetic on those rates.
The call | Rate | This month | Credits |
|---|---|---|---|
Person entity watch, 30-day freshness | 5 per changed person | 4 of the 94 move | 20 |
Person discovery watch, 30-day | 0.5 per new person | 25 new matches | 12.5 |
First run of each watch | Free baseline | 2 baselines | 0 |
Any run with nothing new | Free | the rest of the month | 0 |
Preview before creating | Free | every filter you try | 0 |
The month comes to 32.5 credits. Crustdata credits run from $0.30 to $0.01 each depending on your plan, and your account's own rates are in the pricing docs.
How can you cut the bill?
Watch fewer people more precisely, and stay on the slowest tier that still beats your sales cycle.
Stay on the 30-day tier unless the alert needs a same-week reply. The same move costs 5 credits at 30 days and 150 at one day.
Key the track so promotions stay quiet. Every suppressed promotion is a notification you don't pay for.
Preview every filter before you create it. The count is free and the delivery isn't.
Ask for the field groups you'll actually read. They don't change the price, but a payload nobody reads means a second lookup that does.
Split the list by tier. The 21 champions can sit on a fresher tier than the 40 contacts from lost deals.
Where do I start this week?
Write the brief first, because every other step reads it.
Write the four tiers for one territory, including the cases you never want routed to you.
Preview a discovery watch on one account list and read the count before you create anything.
Create one entity watch on 20 people you would actually act on, with keys set and no channel attached.
Read its baseline run, then hand one real alert to Claude with the brief and see whether you agree with the tier.
Once the tiers come back right on one territory, add the channel and the drafting step.
You can start on a free Crustdata account or book a demo at crustdata.com.
Job change alert FAQ
What is a job change alert?
A job change alert is a notification that a person you track has started a new role. It carries the change itself, the new employer and title, and whatever profile fields you asked for. It says nothing about whether the move matters to you, which is the part this guide covers.
How do you track job changes for a list of contacts?
Create a person entity watch on the list and track experience.employment_details.current with the type added. One watch holds up to 10,000 people, the list stays editable, and you're charged for each person who changes.
How quickly will a job change show up?
Within the freshness tier you set, which runs from 30 days down to one. Checking more often doesn't make detection faster, because the watch compares against the last refresh. The 30-day default costs 5 credits per changed person and the 1-day tier costs 150 credits.
Why does a watch fire when someone gets promoted?
Because a promotion arrives as a new element in the current employment array, exactly as a move does. Add a keys list naming crustdata_company_id and name to the track leaf, and the watch fires only when both are new to it.
Can Claude set up the watch, or do I have to write the call?
Claude can create it. With the Crustdata connector on claude.ai or the plugin in Claude Code, you describe the people, the change and the channel. Claude makes the call. Writing it yourself matters when the watch lives inside a product.
How is this different from a saved search that emails me?
A saved search re-runs and returns everything that matches, including everyone it returned last time. A watch delivers only what's new or changed since the last run and bills on that, so the cost follows movement.
Can I track job changes for everyone in my CRM?
You can, up to 10,000 people per watch, and you pay only for the ones who move. The question is whether you want every move routed to a person. The brief is what keeps a 10,000-person watch from becoming a 10,000-alert channel, because what fires gets logged, and only the tiers you named reach a person.
Do I need a webhook to use this?
No, a pull-only watch works. Send an empty notifications array and read the results from the run endpoints instead, which the connector can call for you. Add a channel once you know which alerts are worth interrupting someone for.
