Can You Build Your Own Buying Signal Tracker With Claude? What Works and What Breaks

Yes, you can build a buying-signal tracker with Claude, and on public sources with a short account list it works well. On LinkedIn it breaks in four places: the daily action limit, page changes that fail without an error, raw signals that are mostly noise, and nobody owning the fix. Build it yourself if one person will own it and your list is a few dozen accounts. If you need LinkedIn engagement signals across a whole market, buy the data layer and keep Claude as the brain.

What a DIY Claude setup does well

Claude is good at reading and drafting. Give it a job post, a company's news page or a prospect's last five posts, and it will tell you what changed, whether it matters for your offer, and write a first line. That reading was the slow part of signal activation. A model can now do it for every record, not only the few a person gets to in a morning.

Public sources with no login are the safe ground. Many applicant tracking systems publish a company's open roles on a public page, so a hiring signal costs almost nothing to read. Add the company's news page, its press releases and your own CRM notes. A weekly job that reads those for 30 accounts and reports what changed is a reasonable weekend build.

We built our own version this way first: six or seven platforms, a lot of time and n8n. Something was always breaking. The trouble was worst where the signals lived on LinkedIn.

Break 1: LinkedIn limits what one account can read

LinkedIn does not tell you the limit. In our experience, a browser agent such as Claude in Chrome, or direct calls from your own logged-in account, gets about 150 actions a day before the platform pushes back, and going past that risks a restriction.

We hit it on 7 October 2026. In one of our test pulls, a read-only job in a logged-in Chrome tab started returning 403 errors after about 150 calls in about 40 minutes.

Run the arithmetic on your own list. Watch 300 accounts and check three pages per account each week, and you need 900 page views, around 130 a day. That is almost the whole allowance before you have sent a message or opened LinkedIn yourself.

An agent in your browser also acts as you. On that same day, a click meant for a reactions counter landed on the Like button, and a like went out from a team member's real account to another company's post. We removed it after about 30 seconds. The notification to that page's admins may already have gone out.

This is why we never use a client's LinkedIn account to read data. We run our own accounts and infrastructure for it, so a restriction lands on us and not on a customer.

Break 2: it fails without an error

Around 26 September 2026 LinkedIn moved member activity pages to a new front end, and we confirmed it on 28 September. The markup that scrapers used to find posts disappeared. Any scraper built on it got zero posts back from profiles that were perfectly healthy. The pages loaded and nothing threw an error. The result just said nothing had happened.

It keeps happening in smaller ways. On 7 October 2026, 13 of about 38 profile loads in our own comment-scanning routine came back as LinkedIn's empty feed, "Nothing to see for now", on profiles that were live. The next day it was 10 of about 50. One reload fixed most of them.

Here is the problem for a DIY setup. A Claude that reads an empty page reports no new signals, which is the same sentence it writes after a quiet week. Nobody sees a failure. The report just gets shorter.

The cheap defence is a canary. Put one account you know was active this week into every run. If the run does not find it, the run failed. Also count the items each source returns, and alert when a source returns zero.

Break 3: most raw signals are noise

In our own daily review of the raw feed, around two thirds of the signals fall into four noise classes. Colleagues congratulating a teammate. Ex-colleagues doing the same on a job-change post. Likes on job ads. Likes on motivational posts that name no product, problem or number. A fifth turned up on 30 September 2026: a recruiter's job ad is filed under the recruiter, so the "hiring company" is a staffing firm posting for a client.

Collecting is the easy part. Filtering is the product. In a run we published on 10 August 2026, 341 of 4,774 raw signals survived, which is 7.1%. At company level, our production run on 21 August 2026 evaluated 161,072 companies and 8,689 passed the ICP filter, about 5%.

Claude can apply those rules if you write them. Give it the lead's company and the post author's company, and it will catch the colleague congratulation. It still costs tokens on every row, and someone has to spot the next noise class and add the rule. Without a hard filter you get more signals and the same pipeline.

Break 4: somebody has to own the fix

Every failure above has the same shape. It happens on an ordinary day, nothing alerts, and the signal is worth less each day it is late. Signals decay, so a fix that waits for the weekend costs you the signal.

Somebody has to notice and fix it the same day. We built and tested our system as a pair: a CTO with a PhD in AI and machine learning, and a co-founder with a decade in marketing, GTM and sales. If you build your own, name the person who does that job before you write the first prompt. Every week that person spends repairing the tracker is a week they are not selling.

Build, buy, or both

If you land in the first two rows, build. The last row is how we run our own outreach too, inside the platform or through Claude Code. Buying signals your AI agent can actually use walks through that loop.

A two-week test before you commit

You do not need to decide today. Run this and let the numbers decide.

1. Pick 20 accounts you would be glad to win.
2. Pick one signal type and two public sources for it. Hiring on the careers page plus the news page is a good start.
3. Have Claude report weekly, with the item count per source and one canary account.
4. Log every fix: what broke, how long it took you to notice, how long it took to repair.
5. After two weeks, count three things: signals that were real, messages sent within three days of a signal, and hours spent on repairs.

If repairs took longer than the reading and writing saved, you have your answer. If your signals live on LinkedIn, scale the calls per day your test used to your real list (calls x list size / 20) and see whether the result passes 150.

Frequently asked questions

Can Claude pull LinkedIn data for me at scale? No. A browser agent or direct calls from your own logged-in account get about 150 actions a day before LinkedIn pushes back. A few dozen accounts fit inside that, and hundreds do not.

Is it safe to run Claude in Chrome on my own LinkedIn account? Only for light reading, and with care: the agent acts as you, and automated access can put the account at risk. Keep it read-only, keep clicks away from Like, Comment and Connect, and stay well under the daily limit.

Can I use Claude with a signal tool instead of replacing one? Yes. Claude does the reading and drafting, and the signal tool supplies filtered, dated events over MCP. A person approves what goes out, as laid out in the complete signal-based selling system.

How do I know my DIY setup is silently failing? Put one account you know was active this week into every run. If the run does not find it, the run failed. Also count the items per source and alert on zero.

When is building it myself the right call? When your list is a few dozen accounts, your sources are public pages, and one person owns the fixes. If one of those three is missing, the maintenance lands on whoever has the least time.