Good morning.
Claude and ChatGPT can connect straight to your email platform, read every send you've ever made, and do almost everything on your list except press send.
We've been running it on one of our autonomous business projects, and the first afternoon I read that project's data, I got three things confidently wrong.
Today's the full playbook: how to connect it safely, the one file it reads before every job, the twenty-minute weekly run that picks, writes and loads your next email, and the quarterly check that stops you acting on a number that only looks like an answer.
— Sam
IN TODAY’S ISSUE 🤖
Everything the AI handles before you press send
Connecting Claude or ChatGPT safely
The list file it reads before every job
The Monday check against your own normal
Picking, writing and loading your next email
The resend that reaches people who missed it
The monthly jobs and the quarterly audit
The three wrong answers from one afternoon
What never to hand over
Let’s get into it.
The AI Does Everything Up To The Send Button
The AI reads your whole account, decides what to look at, writes the next email, builds the segments, and loads a finished draft into your platform with the audience set. Then you preview it and press send.
Most platforms already stop a connected AI at the send button:
beehiiv doesn't let it send or schedule at all.
Kit makes you confirm inside Kit before anything goes out.
ActiveCampaign's connection has no send tool.
Klaviyo, and Mailchimp through its API, will send if you let them, so the setup below blocks that.
It's the right place to stop anyway. The send is the one step you can't take back, and it takes you thirty seconds.
When | What the AI does | Your time |
|---|---|---|
Once | Builds the list file it reads before every job | About an hour |
Every Monday | Checks last week against your normal, briefs the next email, drafts it and loads it | About 20 minutes |
After a send that matters | Resends to the people who didn't open | 5 minutes |
First Monday of the month | Refreshes your segments, checks your welcome sequence and cleanup flow | About 20 minutes |
Once a quarter | Audits the whole archive | About an hour |
The smallest version takes ten minutes: connect, skip the list file, and run the Monday check once with one extra line telling it to work out your normal from your last ten sends. It only reads, and it tells you whether last week was ordinary for your list.
Connecting Claude Or ChatGPT
Most email platforms now have an official connector, which lets Claude query and update your account directly instead of you exporting a CSV and hoping you picked the right columns:
beehiiv, Kit, Klaviyo and ActiveCampaign all have one.
Mailchimp's official Claude connector is built for planning campaigns, so past stats and drafts go through its API.
All of this works in ChatGPT Work too. Klaviyo, Mailchimp and ActiveCampaign have ChatGPT apps, and beehiiv and Kit connect as custom connectors on business workspaces. Every prompt below works in either. I say Claude throughout because that's where we run it.
If you're on Kit, pull your archive before October 15. From that date Kit's API only returns email stats from the last five years, and older data takes a request to their support team.
Set the permissions before the first job. In your connector settings, allow reading and drafting. Block anything that sends, schedules, deletes, unsubscribes or suppresses. Then put the same rules in writing, because the AI reads them before every job:
You may read everything on my account, write drafts, build segments and tags, and update list-file.md. You never send, schedule, delete, unsubscribe, suppress or import subscribers, and you never edit an automation that is live. If a job needs one of those, stop and tell me exactly what to do in the app.Run two checks before you trust any number.
Check 1: confirm what your platform means by "click rate." Some platforms divide clicks by delivered emails and some divide by opens, and the two answers differ by roughly a factor of two. The platform on our project divides by opens, even though its own help page describes it the other way.
To check yours, pick one post, count the clicks and opens yourself, and compare them to the reported figure.
Check 2: ask for the field list before you ask for a conclusion. It takes two minutes, and it tells you what you're able to ask:
Connect to my email platform and pull the complete stats payload for my most recent sent email. Show me every field name the response contains, with its value, and do not summarize or interpret any of it. I want to see what the platform exposes before I ask you a question about it.The List File
The list file is one document the AI reads before every job, so on Monday morning it already knows your list and you don't have to explain it again.
It holds two kinds of thing.
What the AI measures itself:
Your normal range for opens, clicks and unsubscribes
The segments on your account
When you send
Your best subject lines
What only you can tell it:
What you sell
What counts as a conversion, and where it's recorded
What's coming up in the next 90 days
How you write
What it must never do
Have the AI build it:
Read my email platform's archive for the last 12 months and draft my list file with these headings: My List, What I Sell And What Counts As A Conversion, Audiences I Send To, Cadence And Calendar, My Normal, Voice, Subject Lines That Worked, What You May Do / What You Never Do, Things To Watch, Change Log. Fill in everything you can measure yourself: the segments that exist on my account, my send days and times from my history, My Normal (median and range of open rate, click rate and unsubscribes over my last 10 full-list sends, and my weekly sign-ups over the last 8 weeks, stating which denominator my platform uses for click rate), and my top 10 subject lines by opens and by clicks. Then ask me, one question at a time, only what you can't measure: what I sell and at what price, what counts as a conversion and where each one is recorded, my launches and offers for the next 90 days, my voice rules or voice file, and anything you must never do. Save the finished file as list-file.md.Why one question at a time. A list of ten questions gets ten short answers. One at a time gets the detail, and the AI can push back when your answer doesn't match what it measured.
Two things to get right:
Where each conversion is recorded. If clients book through an application form or a call, your email platform never sees it. Every "what converted" question will undercount your biggest revenue line unless the file says so.
Your voice. If you already have a voice file, point to it. If you don't, paste two past emails you liked. That's enough to start, and the list file gets better every time you edit a draft.
Where to keep it:
In Cowork, save it in the folder Claude works in, so every job can open it.
In ChatGPT Work, add it as a project file, and paste the permissions block above into the project's instructions.
Every job below runs from the prompts in this issue. If you'd rather have it installed, Cortex members get the whole system as a download:
A Claude Skill that builds your list file and runs each job, the same setup for ChatGPT Work, and four filled-in list files to copy from. Get the Email List Operator Build Pack →
Once paid, come back to this issue and refresh (on the web), you’ll see the link below.
The Weekly Run
Four steps on a Monday, in order. About twenty minutes of your time, most of it reading and deciding.
1. The Monday Check
Was last week normal for your list?
Read list-file.md. Pull every send from the last 7 days and my subscriber changes for the same week. Compare each send against My Normal (open rate, click rate, unsubscribes and spam complaints) and say whether each number is inside or outside the range. Compare the week's sign-ups, unsubscribes and net change against the 8-week average and range. Label every number measured or inferred. Keep it to one short table and at most three flags, then end with one line: anything to act on this week, yes or no, and what."Normal" here is the range across your own last ten full-list sends, from the second-lowest to the second-highest.
An industry benchmark tells you how other people's lists behave. Your own range tells you whether this week was unusual for you, which is the only question a Monday check needs to answer. Most weeks the answer is no, and that's a result.
The three-flag limit is there because without it the check reports every small movement as a finding, and within a month you stop reading it.
2. The Next-Send Brief
What should the next email be?
Read list-file.md. Pull my full-list sends from the last 90 days sorted by click rate, and read the reader replies, poll results or survey answers I paste below. Check Cadence And Calendar for anything due in the next two weeks. Give me three options for my next email. For each one: the angle in one sentence, the one thing I want readers to do, the audience, and why now, citing the sends or replies that support it. Recommend one and say why.Paste your replies under it. Email platforms can't read your inbox, so this is the one input you bring, and it's usually the best one: the question three readers asked last week is a better next email than anything the numbers suggest.
Why sort by clicks: opens tell you a subject line worked, and clicks tell you the email gave people something to do, which is what you want more of.
3. Draft And Load
Write it, and put it in the platform ready to send.
Read list-file.md and write the email for the option I picked, following the Voice section exactly. Write five subject lines, each with a preview line, and compare them with Subject Lines That Worked: say which of mine each one is closest to and flag any that promise more than the email delivers. Put your pick first. Then load the email into my platform as a draft, never scheduled and never sent, with the audience from the brief and every exclusion listed under Audiences I Send To. Read the draft back from the platform and give me the pre-send checklist filled in: audience name and recipient count against my list size, exclusions present, subject and preview as loaded, every link, and the send time you'd suggest in my time zone.Always have it read the draft back. What you asked the platform to save and what it saved can differ, especially when a draft starts as a copy of an old email. On one of our projects we found two ways copies go wrong:
A copied email kept the old email's subject line.
Another lost the segment that stops existing buyers getting a sales email.
The read-back catches both before anything goes out. Then it's your turn, in the app:
Open the draft and preview it.
Click every link.
Check the audience count against the checklist.
Press send.
Platform notes:
ActiveCampaign can't create a campaign through its connection, so create an empty one with the audience set and have the AI fill in the message.
Mailchimp drafting goes through its API.
4. The Resend
A second chance for the people who didn't open.
Use it only for sends that matter: a launch, an offer, a deadline. Skip it on routine issues, and resend any email once at most.
It's been two days since [the send]. If Resends allowed for in list-file.md covers this send and it hasn't been resent, build a segment of everyone it was delivered to who didn't open, excluding anyone who has unsubscribed since, and tell me the count. Load a copy as a draft to that segment with the runner-up subject line and a new preview, and set the subject, preview and audience on the copy yourself rather than trusting what it inherited. If the send had a deadline that has passed, skip it and tell me why.People who didn't open never saw the email, so for them the resend is the first time they see it, under a subject line they haven't already passed over. On one of our projects, the resend to non-openers produced the best click-to-open rate of a whole launch.
The Monthly Jobs
First Monday of the month, straight after the Monday check. About twenty minutes.
Read list-file.md. Refresh four segments and report each count against last month: engaged (opened in the last 60 days), cold (no opens in 90 days or more, excluding anyone who joined in the last 90 days), buyers, and new (joined in the last 30 days). Then report my welcome sequence email by email: open rate, click rate and unsubscribes against My Normal, and name the email where people drop off. Check that my cleanup automation is running and how many people it removed this month. Refresh My Normal, add a line to Change Log, and list anything in Cadence And Calendar that has ended or is coming up in the next 60 days.What each segment is for:
Engaged is who your next send is mostly for.
Cold is who your cleanup flow will remove next. It leaves out recent joiners so it doesn't grow every time your list does.
Buyers get excluded from anything selling what they already own.
New is who your welcome sequence is working on. Most sequences have one email where people stop reading. The AI finds it, and rewriting it is your call.
The job also refreshes your normal. A normal measured in March is the wrong yardstick by October, and a Monday check against a stale one flags things that are ordinary now.
The Quarterly Audit
Once a quarter, the AI reads your whole archive, and this is where the three wrong answers came from.
The open rate on one of our autonomous business projects had been drifting down for months. We run a number of these, they all send email, and this pattern turns up in more than one of them. This project's sends opened around 51% at the start of last year and around 46% by this summer.
We read that the way anyone would: the content was getting less interesting, the list was getting tired, and something needed to change. We built a whole quarter's plan around it.
Then I read the archive properly, and almost every part of that story fell apart:
The list was growing. New subscribers were arriving at roughly four times the rate they had a year earlier.
Most of the leaving was ours. A cleanup automation we'd set up ourselves was removing people who had stopped opening.
The new subscribers read fine. Our next explanation was that all those new people were dragging the average down. Measured directly, they read about as well as everyone else.
Your dashboard reports one open rate with several movements folded inside it. The first three reads pull them apart, and the fourth checks what the list earns:
The Mix separates the kinds of email you sent.
The Flow separates who arrived from who you removed.
The Cohort measures how the new people read.
The Money finds what converted, and when.
"Your content is getting worse" and "your list is growing fast" can produce the same number. Run the reads in this order, because each one changes what the next one means.
The Mix
Which sends are you averaging together?
Your archive holds four kinds of post, and your platform averages all of them into one open rate:
Full-list sends
Segmented promos
Paid-subscriber or buyer-only emails
Posts that never mailed
When we sorted that project's archive, about a third of the posts were something other than a full-list send. Averaged together, the delivered list looked like it had dropped about 4% over the summer. Isolating full-list sends, it had held flat within 1%.
Pull every published post from my archive since [DATE] with its send stats. Classify each one into: full-list newsletter sends, segmented sends under [N] recipients, paid-subscriber-only sends, and posts that never mailed. Show me the count in each group and the recipient range within each. Then report the monthly average open rate, click rate and recipient count for the full-list group ONLY, as a table by month. Do not blend the groups, and tell me explicitly if any post is hard to classify rather than assigning it without saying so.We use about 80% of the list as the full-list threshold. Get it wrong and every trend below inherits the error.
The Flow
Who arrived, who left, and how much of the leaving did you cause?
Read straight, that project's last three months said it was running to stand still: 91 subscribers leaving for every 100 arriving.
Then I checked its automations:
A cleanup flow removing people after six months without an open accounted for 71% of all the leaving.
Organic churn was 27 per 100 arriving.
Taking the cleanup out, the list had grown about 13% in the quarter.
For the last 3 months and the last 12 months, report: new subscribers, churned subscribers, net change, and the acquisition source breakdown with counts. Then list every automation on my account and tell me which ones remove, unsubscribe or deactivate subscribers. For each of those, pull its stats and report how many subscribers it removed in the same window. Subtract that from total churn and show me organic churn separately, as a count and as a rate per 100 acquired.Watch for counts that are identical in both windows. If a source shows the same number over three months and twelve, nothing came from it before three months ago. On our project, that dated the whole ad program a month away from where we'd have put it from memory.
The Cohort
Do your new subscribers read like your old ones?
Build two segments: subscribers who came through paid media, and everyone who joined before the ads started.
Ours came back with the paid cohort opening about twelve points below the base, which looked like a serious problem.
Then I checked what period each number covered. They were lifetime figures, and the base's rate was spread across years of sending, including years when the whole list opened better. Against the project's recent monthly rates, the same paid cohort sat about two points under the house norm.
Build two segments. Segment A: active subscribers whose acquisition medium is paid. Segment B: active subscribers who joined before [DATE the ads started]. Report each segment's member count, open rate, click-through rate, and the count and percentage on a paid tier.
Then tell me the time period each of those rates covers. If they are lifetime figures, say so, and compare the paid cohort against my recent monthly full-list rates from the Mix read rather than against the base's lifetime rate. Flag explicitly if the two are not comparable.Without the prompt's second paragraph, you get two numbers side by side that look like a fair comparison and aren't. Hand "our paid subscribers open twelve points below our base" to whoever runs your ads, and they'll pause the ads for a reason that isn't real.
The Money
What converted, and when?
Year over year, that project's conversions fell by about 70%. Then we looked at where the earlier year's conversions came from:
One send produced a third of the year on its own.
The email that closed the same launch five days later produced another tenth.
The rest converted at roughly this year's rate.
So the question changed from "why did conversion collapse?" to "when did we stop running launches?", which has a cheaper answer.
For every post in the full-list group, report the conversion events attributed to it (upgrades, purchases, or whatever my platform tracks). Sort by count. Then show me the distribution: what share of total conversions came from the top 3 posts, and what the median post produced. Tell me whether conversions are spread across regular sends or concentrated in a small number, and identify what those top posts have in common in format or timing.The Check That Makes All Of This Trustworthy
I reached three confident conclusions that afternoon, and all three were wrong in the same way:
Acquisition. I netted unsubscribes against the change in delivered recipients and concluded the project was barely acquiring anyone. It was acquiring at more than ten times the rate I'd calculated.
Churn. I concluded churn was eating 91 of every 100 new subscribers. That was before I subtracted our own automation.
The paid cohort. I estimated its open rate from list arithmetic and built an explanation of the whole decline on it. Measured directly, the real figure was thirteen points higher.
All three came from arithmetic over aggregates, and asking the platform for the thing measured directly corrected each one. The third took two segments and about forty-five seconds, after three hours of arithmetic.
So the rule that runs through every prompt in this issue: before a number becomes a decision, find the thing that measures it directly and ask for that instead. Add this line to any analysis you run:
For every conclusion you give me, state whether it is measured or inferred. If measured, name the field and the population it covers. If inferred, show the calculation and name the single assumption most likely to make it wrong, then tell me what I could query to check it directly instead.What Never To Hand Over
Sending and scheduling. The AI loads drafts and you press send. On Klaviyo, and on Mailchimp through its API, block the send tools in your connector settings, because both will send with no confirmation.
Deleting, suppressing and unsubscribing people. Your cleanup automation does this on rules you set once. Nothing else should, and on some platforms deleting someone also erases their history from your old reports.
Imports. A bad import can add people who never signed up, and you're the one answering for it.
Live automations. Edits to a live flow can go out to people already in it. Have the AI draft the change, and publish it yourself.
Scheduled runs with write access. Running the Monday check automatically is useful, and Cowork and ChatGPT can both do it. Give the scheduled task a read-only connection, because nobody is there to approve anything when it runs.
Copies of old emails, unchecked. If a draft starts life as a copy of an older send, reset the subject line, preview and audience on the copy, and check the exclusions are still there.
What This Looks Like By Business Type
Consultant or coach
Keep launch emails out of your normal, or every launch week will look like a crisis.
Put where clients book (an application form, a call) in the list file, because your platform can't see it.
Use the resend on launch emails, which is where it pays most.
Read list growth over twelve months before three, since launch lists grow in spikes and shed between them.
Agency
Your conversion is a booked call, so the next-send brief should draw on what prospects ask on calls as much as on the numbers.
Move client-only updates to their own list, or they'll drag down every average.
On ActiveCampaign, create the empty campaign each week and let the AI fill it.
Creator or publisher
The Monday check is worth the most here, because you send often enough for a normal to mean something within a month.
Build the Cohort read per acquisition source, since subscribers from a lead magnet, a recommendation and a paid ad read differently.
Test every subject line against your own past winners.
E-commerce
Start the brief from the calendar and what's in stock.
On Klaviyo, block the send and suppress tools before anything else.
Give the monthly job more time, because your flows (welcome, abandoned cart, post-purchase) send more email than your campaigns do.
Set your attribution window before the Money read.
Want The Build Pack?
You can run everything above from the prompts in this issue. If you'd rather have it installed, Cortex subscribers get the full Email List Operator Build Pack as a download:
The installable Claude Skill that runs the whole system: builds your list file, runs the Monday check, drafts and loads your next email, handles the resend, the monthly jobs and the quarterly audit
The same system for ChatGPT Work: project instructions and every prompt, ready to paste
The list file template and four filled examples (a consultant, an agency, a creator newsletter and an e-commerce brand)
Setup guides for beehiiv, Kit, Klaviyo, Mailchimp and ActiveCampaign: what the AI can and can't do on each, and how to limit it to reading and drafting
The pre-send checklist, segment definitions for every job on every platform, and scheduling guides for the Monday check
The tune pack: eight ways this drifts week to week, and the fix for each
An example week and two example quarterly audits, so you know what good output looks like
This is only available to paying Cortex subscribers.
What is Cortex?
Cortex is the premium monthly membership version of Bionic Business, helping entrepreneurs and business owners grow and scale with AI and Agents.
In addition to 3-4 full, unrestricted regular issues per month, you also get:
1x SIGNALS Strategic Briefing issue (strategic moves you must make now to secure your business revenue, market share, and profits).
1x CIRCUITS Tactical Guide issue (workflows, how-to’s, prompts, very practical implementation inside a self-driving business).
Access to all Agents and Skills issues (full details, download files, tutorials).
Claude Cowork as your marketing team special issue (full guide on how to run Cowork as an agent marketing team for your business).
50% OFF workshops on prompt engineering, agent operating systems, and more. Only active Cortex subscribers get any kind of discount, no one else.
So, if you’d like to receive full issues and the monthly special Signals and Circuits issues, then you should sign up for Cortex, which is $50/month (and you have full archives of previous regular issues for as long as you’re a paying member).
One note on Signals and Circuits: You only get these special issues for the months you’re subscribed—which means if you skip a ~30 day period, about a month, you won’t get them for that month. There are no back-issues for past months, so now is your chance to lock-in these upcoming issues while you can.
After you’ve upgraded to Cortex, log in to the website and come back to this issue, and the Build Pack will be unlocked below.
Ready? Get the Build Pack, full regular issues, and the upcoming Signals and Circuits issues:
Cortex is a monthly premium version of Bionic Business, helping entrepreneurs grow and scale with AI and Agents. You can cancel at any time. But there are no refunds, for any reason. All sales are final.
One Question Before You Go
One tap is enough, and it decides which formats get more of the slots this quarter. If you want to say more, the box after your vote comes straight to me.
Which kind of issue do you get the most from?
The first afternoon I pointed Claude at that project's list, three of my readings turned out to be backwards.
A weekly check against the list's own normal would have caught every one of them long before I sat down to look.
Start with the Monday check this week. It takes ten minutes and only reads, and by the third Monday you'll know what normal looks like on your list, which every other job in this playbook depends on.
Talk soon,
Sam Woods
The Editor
P.S. The Monday list check in this issue is one of the twenty jobs that run without you by day 60 of The 1 Person, $1M Business. Enrollment is open to Cortex members until tomorrow, Wednesday, October 7, at 11:59pm ET. Not a member yet? Join Cortex to get in.
.
