A researcher that never sleeps. It reads the newsletters, watches the YouTube channels, checks X and the official release notes, files the best of it in one place every day, and pings your phone the moment something urgent drops. Running cost: about $10 a month.
Write a job description in plain English: what to hunt, what good looks like, where to file it. Give it one tool and one schedule. The simple version takes twenty minutes. Both recipes are below.
The noise problem
To make this newsletter worth your inbox, we need to know what is actually happening in AI for business. Not the hype. The releases that change what is possible, the pricing moves that change what is sensible, and the operators quietly sharing what is winning them customers and growing their revenue.
Keeping up properly means reading the newsletters, watching the YouTube channels, trawling X and checking the official release notes. Every day. Between the two of us that was heading for three hours a day, call it 20 hours a week, spent filtering instead of building.
So we each hired a researcher. Neither of them is human.
Two builds, one job
Here is the part we did not plan. We built our researchers separately and ended up with two completely different employees doing the same job. Which means you get two recipes, and you can pick the one that fits how technical you want to be.
Built in Claude Cowork as a scheduled task. About twenty minutes to set up. The whole agent is one instruction document written in plain English: what to hunt, what good looks like, where to file it.
It runs daily on his machine, scanning the newsletters he already subscribes to and then sweeping the web for primary sources. If his laptop is closed it simply catches up when he opens it.
It works when the office is open.
Built with Claude Code as a scheduled routine running on Anthropic's infrastructure. Under an hour for the first version. The job description is a handful of plain English files: a mission, a quality bar, an operating procedure. 247 lines of English. Zero lines of code.
It runs three times a day whether our laptops exist or not, and one extra ability, wired in through Apify, lets it transcribe YouTube videos and read X. It reviews the full transcript of every new video on the channels we track, not just the title.
It works while we sleep.
What they do all day
Both researchers hunt the same two things: AI releases and advancements that matter if you run a business, and real operators showing how they are using AI to win customers and grow revenue.
Everything they find is scored out of ten against a scoring system we wrote for them, and the best items are filed into one shared workspace with a summary, why it matters and the source. Each checks the other's recent filing first, so we never get the same story twice. And anything urgent, a major release or a market move, skips the queue and pings our phones immediately.
A few of the catches from their first week:
- Claude Cowork going 24/7 and arriving on mobile.
- Attio's latest raise, positioning it as the AI-native answer to HubSpot and Salesforce.
- Corey Haines shipping a 46-skill AI marketing pack.
- Marketers publishing their actual frameworks for handing AI real marketing work.
What broke (the useful bit)
Week one was not clean, and the failures taught us more than the setup did.
The cloud researcher blew its monthly scraping budget in two days. Reading X turns out to be the expensive habit; YouTube transcripts cost fractions of a cent. The fix was rationing: X once a day, capped per account, transcripts unlimited. Know where the costs hide before you put anything on a schedule.
It also once filed its daily report as an email draft instead of sending us a notification. The fix was one new sentence in its job description. That is the strange, brilliant part of managing AI employees: you do not debug them, you amend their job description.
One job. One tool. One schedule. One place to file the results. That is the whole recipe.
Which one should you copy?
Both, in order. Start with the desktop employee: one Cowork task, twenty minutes, nothing technical. Graduate to the cloud version when you want it running with your laptop shut, a history of every change you make to it, or a copy doing a second job.
And a researcher is simply the hire that fits a media business like ours. The recipe works for any job you can describe in plain English. Here are six first hires that generate business, not just information:
Builds a shortlist of companies that look exactly like your best customers, with the reason each one fits.
A one-page brief on the person, the company and what has changed since you last spoke, ready before every call.
Sweeps your pipeline, fills the missing fields from recent emails and notes, and flags the deals going quiet.
Watches competitor pricing, launches and reviews, and tells you the day something moves.
Reads the week's conversations and drafts every follow-up you did not get round to sending.
Reads the overnight inbox and hands you a short list of what actually needs you, and what can wait.
The researcher is live today. Several of the others are next on our own hiring list, and we will publish each job description as we write it.
Try this week
Find the most tedious 30 minutes you spent this week. Write the job description for it in plain English: what to do, what good looks like, where to put the result. You now have your first AI employee's contract. Hiring it takes twenty minutes.
Running costs, for the sceptics: nothing beyond the Claude subscription we already pay, plus about $10 a month in API fees.
- Pick one tedious job you repeat every week.
- Write its job description in plain English: what to do, what good looks like, where to file the result.
- Give it one tool and one schedule. Cowork for the twenty-minute version; Claude Code when you want it running with your laptop shut.
- Point the output somewhere you already look, and when it gets something wrong, amend the job description.
Jared and Kieran
The AI Natives