This one covers October 20 through 27. Most of the week went into one idea: every AI tool I use should know the same things about my business, and none of them should ever be told what doesn’t belong there.

Friday (my AI chief of staff) - One memory for every AI tool I use

Every time I opened a different AI tool, I started from zero. Who my clients are and what I decided last week. I’d explain it again, and again, and it still got things wrong.

So Friday (my chief of staff) now keeps one shared brain: a private, version-controlled set of plain pages about my work and the decisions behind it. Session notes and voice captures go in one end. A daily job distills them into pages. A sync job pushes the pages out every 5 minutes, so any AI tool I use reads the same thing.

The part I care about most is what stays out. Nothing reaches the brain without passing a leak test. It blocks passwords and keys, phone numbers, street addresses, money amounts, customer names, and a list of topics I keep out of anything shared. A note that trips the test gets quarantined. It doesn’t get filed.

The final review caught something I wouldn’t have. The pages cited their sources by file name, and the newest file name carried a customer’s last name. We rebuilt the citations to use project, type and date before anything went live, and the name check now covers last names too.

It went from design to live in one day. I started Sunday around 6 AM, and it was live at 4:36 PM: 16 build tasks, 44 commits, 3,517 tests passing at merge, and 3 background jobs running.

The gap: the redaction sometimes eats more than it should. One summary line lost a chunk of words to a redaction tag, and that’s on the fix list.

Before you give an AI more of your business, decide what it will never be told, and test for it. The test caught a customer’s last name in a file name, the last place I’d have looked.

Friday - Saving a note by voice from the truck

The next step was the phone. One tap saves a note into a shared folder. The sync job pulls it in, runs the leak test, and files it into the brain. The same folder holds a read-only copy of the brain pages, so an AI on my phone can answer from them.

I dictated a note while driving, and it was filed at the next morning’s run. The morning brief now tells me how many phone notes came in.

The first real note I dropped in got set aside. It had details in it that don’t belong in a shared brain. A cleaned-up copy was filed instead, and the full version stayed on my machine. That’s the guard doing its job on day one.

Early phone tests turned up a sneaky one. When the phone AI’s own built-in memory was out of date, it said “brain loaded” while reading the wrong thing. The instructions now say the shared folder is the only source.

This one went from spec to merged in about two and a half hours on Monday morning and was live at 10:07 AM. 3,764 tests passing. I planted 31 deliberate bugs to see if the tests would catch them, and they caught 29. The other 2 didn’t change what the code does. Review closed 8 findings before merge.

Your best ideas show up in the truck, not at a desk. If capture is hard, they die there. If capture can leak, you stop using it.

Friday - My agents check the meter before they start work

Friday runs builds in the background, and each one uses part of my weekly AI plan. Last week that was a rule Friday followed each hour. Now it’s a coded gate that reads my live usage before any build starts.

During business hours, work only starts if I’m at or under the pace for the week. At night and on weekends, it starts only if the week’s usage plus that build’s expected cost stays under a hard stop at 85% of the weekly limit. Builds are tagged small, medium or large, and the expected cost of each size is learned from the last 20 clean builds.

Review fixed three things before it went live. An idle window read as “not collected” instead of zero. A learned cost could round down to nothing and wave everything through, so it has a floor now. And the picker now asks the gate which sizes fit before it chooses a card, so it stops picking work that can’t run. If the usage reading can’t be fetched, the gate falls back to its own ledger and stops.

Agents that work while you sleep will burn whatever you give them. Give them a budget that reads the real meter.

Friday - Build time on my calendar, and a task board that tells the truth

Two small jobs, both running on their own.

Every weekday morning, one finds gaps of an hour or more on my work calendar and blocks them as build time, with 15 minutes of buffer on each side. It only creates and deletes its own blocks, and prospects can still book the open slots around them. The first review found 2 serious problems before it ever ran. A block I’d moved or edited could still count as its own, so deleting it could have sent cancellations. And started mid-afternoon, it would have planned the whole day. Both got fixed first. 104 tests, every rule checked with a planted bug.

The other is a daily report on my task board: cards past their date, cards that have sat untouched (a week in progress, three weeks in to-do), and cards waiting on me. A second check flags anything marked complete whose work isn’t actually done. It only reports. It doesn’t touch a card.

Most of an owner’s time isn’t stolen. It just never gets protected.

AI Marketing Department - The marketing engine: posts as long as the answer

My marketing engine shipped 4 releases in 7 days. The one I’d point a business owner to: post length now follows what the reader searched for.

It used to size every blog at about 1,300 to 1,700 words. Now it picks a length per post. A quick answer gets 600 to 900 words. A how-to gets 1,000 to 1,400, a full guide stays at 1,300 to 1,700, and a comparison gets 1,600 to 2,200. The quality check uses the same limits, so a short post doesn’t get flagged as thin. A post with unclear intent still defaults to guide length.

Also this week:

• A work-completed report for ad accounts. The audit finds the problems. This report records what got fixed, and every change has to name the finding it closes, what it should improve, and how we’ll see it working. A change that can’t say that gets rejected as activity.
• No duplicate posts, on every path. Last week’s guard covered one publishing path. This week’s bug sweep found the same gap in another one. Now a marker on disk blocks a blind re-send after any uncertain publish, and a test locks it.
• One bad row can’t stop the scheduler. A broken line in the job list skips itself, says which line it was, and lets the rest run.
• The caption check from last week caught three more wordings, and teammates’ changes now get line-by-line review before merge.

The bug sweep confirmed 5 findings, and all 5 were fixed for the whole class of bug, not just the line it reported. The engine’s tests went from 3,285 to 3,476, all passing.

Google doesn’t rank on length. A short local question deserves a short answer, and nobody should pay for padding.

AI Builders Labs - A lead engine built to never message someone who said no

For AI Builders Labs, I built a lead store and two outreach skills, one for email and one for LinkedIn. Every lead gets a fit check. A batch of fewer than 20 verified leads gets held instead of sent. Eight kinds of reply each have a set next step. An opt-out on one channel stops that person on every channel. A database rule blocks any send to someone marked do-not-contact, and a cap of 20 sends a day lives in the code, not in a prompt.

Review found one serious gap before a single send. If one person showed up on two rows, once by email and once by LinkedIn, an opt-out recorded on one row didn’t cover the other. The fix ties the opt-out to the person, not the row, and it has to land before anything goes out.

94 tests, plus 5 planted bugs, all caught. It hasn’t sent anything yet. Choosing the sending mailbox comes first.

If a rule protects your reputation, put it in code. A prompt is only a suggestion.

AI Builders Labs - The agency AI scorecard is live

Last week’s build log said the 10-question scorecard for agency owners was built and waiting. It went live October 20. An agency owner answers 10 questions, gets a score in one of 3 bands, and sees the first fix to make. Their answers trigger a breakdown email, and a second automation tags them with their band and swaps the tag if they retake it.

The domain I wanted to send from was already in use for other mail, and moving it would have broken that. So the emails go out from their own subdomain, set up and verified before the first test. The first end-to-end test passed the same afternoon. Now I’m watching for the first real submission.

Friday - My weekly scorecard fills itself in

I track six numbers a week on a leadership scorecard. Every Monday at 6 AM, a job now pulls them from the sales scorecard and adds the new week’s column. It never edits an existing cell. If a write fails partway, it rolls the new column back, and it won’t write a blank column.

The bug sweep is what found that a partial failure could leave an empty column behind. The first live run was one I watched: column written, read back and checked, and a second dry run changed nothing. Then the Monday run got switched on.

The limit: it copies whatever the sales sheet says. Fix a typo there later, and the old cell stays wrong until someone fixes it by hand.

A number someone has to type in by hand shows up late, and a late number can’t steer anything.

Personal inbox agent - In progress: a second agent for my personal inbox

I added a second agent, Edith, for my personal inbox. Every 2 hours she sorts it into Todo, Delegate, Follow Up, or archive. She never replies, clicks a link, pays anything, or follows instructions in an email. She pings me only for a new Todo or a priority sender, sends one recap at 5 PM, and stays quiet from 9 PM to 7 AM. Anything that’s business gets labeled for Friday, and Friday picks that label up every morning.

It’s a one-week trial. She can see my personal email, not my work email, and her runs count against the same AI plan, so I’m watching the weekly usage before deciding whether she stays.

One agent doing everything gets confused about whose job is whose. Two agents with clear lanes hand work to each other instead of piling it on you.