My owner's queue looked empty for 4 days. It wasn't. A task someone had flagged for sign-off was sitting there hidden, and the thing hiding it was my own system second-guessing the person who flagged it.
The person who wrote the task had marked it for owner review. The risk check scored it as nothing and hid it anyway. I changed it so anything high risk that's explicitly flagged for the owner always lands in their queue, whatever the automatic check thinks.
Most people see a quiet error log and figure nothing's going wrong. Mine wasn't broken. Almost nothing in my system reported failures to it anymore, so real problems could happen with zero trace. 89% of everything it ever caught came in its first 8 days.
That was out of 3,336 records total. By July it was catching about one a week. The last thing still feeding it got turned off on purpose on August 30.
Most people leave the unfinished thing up and figure nobody will find it. I take it down all the way. Half-live is worse than gone.
I had a homework form built for customers, and I rejected it. The page, where it lived, and the public way in were all removed and checked gone. The part that collects the answers stayed put until I review it top to bottom.
Almost nine years at Keller Williams. My big takeaway: the brokerage doesn't matter.
I lived with that for six of those nine years.
What changed it was realizing that after nine years, they didn't know who I was, didn't know my goals, and their outreach was blast texts and event emails.
A personal brand cannot carry the culture of a company.
Starting today with Homes of Idaho. When I told them the brokerage didn't matter, the broker looked me in the eyes and said, "My life mission is now to prove that statement wrong."
You might be at the wrong place, comfortable enough that you never go looking. I know that was me.
Owners using my assistant shouldn't feel a thing while I build what's next. So the prep for anything new happens where it can't touch a single live account, and going live stays its own separate job.
Two pieces of groundwork for how new owners get set up went in that way. Before either one went in, I compared them line for line against what owners were actually running. Nothing they used changed.
I closed bugs without writing a single fix. All it took was one habit I keep before I touch any reported problem: I check whether it's still a problem.
Two open bugs got closed after a look at what was live. A comment reply bug that used the wrong name had been fixed three days earlier. Comments from one customer that weren't reaching my system were flowing in again as of the day before.
My last 8 LinkedIn posts sounded robotic and copy-pasted. Same template, same phrases, every one. A post should sound like me talking and still keep the real technical detail.
I built a step that rewrites every post in my voice and leaves the technical facts alone. I ran it on 44 of the 45 posts already written, and kept every original backed up in case anything went wrong.
Most agents onboard a client by chasing them for a bio and a photo. I'd rather they tell my system about their business themselves, in one simple form, and be done.
I built a first draft: name, bio, specialties, how they want to sound, and an optional headshot, behind a private link that only works once. I reused a secure form I'd already proven in another product instead of starting from scratch. I tested it live in a browser, from submit to everything landing where it should. The questions it asks still weren't final.
My assistant was missing some people's timelines, and the fast patch would have gone stale. So I fixed the cause instead. A patch that sets up the next bug isn't a fix, even when the real fix is more work.
My assistant was missing some people's timelines. The quick fix was to fill in a date once and leave it there. That date would have gone stale. So I rebuilt how it reads the way people say when, so it works it out fresh every time. I checked it against every real phrasing I had. Nobody got moved up sooner than they meant.
I went through every way my system could reach an owner and found some that would act for anyone who asked. They'd been like that for months. Now nothing texts an owner or sets things moving on their side until it proves it came from me.
I locked the doors that could reach an owner directly: the one that takes their incoming messages, the one that hands a conversation to them, and the line that texts them. Each one got tested live. Requests without the secret got turned away, the right ones got through, and ongoing checks confirmed normal messages still arrived like they always had.
Any automated task my system starts from another task has to name its parent. Every time. That's how I trace exactly which task spawned which.
And I proved it worked on a live run before enforcing it.
I added the rule, then watched a real helper task on the live system record its parent correctly. I still held off on enforcing it. Two related gaps were open, and I wanted those closed first.
People underestimate how simple things need to be for someone to say yes.
If your offer doesn't make sense in a five minute conversation, no amount of research is going to make the selling team comfortable.
That's a hill I'll die on.
An old report said my assistant skipped lead questions 29% of the time. I measured it again myself, and the real number changed what I did next.
I went back through real conversations. It happened about 2% of the time. I tried one fix and it showed no real effect, so I didn't ship a behavior change.
When someone texts STOP, I block that phone number on every contact record in the account. Even before I know those records are the same person.
And if I can't tell who sent a message, I treat them as blocked.
My assistant checks the opt-out against the phone number or email itself, not against who it thinks the person is. Any contact sharing that number is blocked. Anyone it can't identify stays blocked by default.
A redrafted text was asking for an email that was already sitting on the person's card. That's backwards. My assistant should know what's on file before it writes, and finish the whole ask in one go.
Redrafted texts were asking for an email or phone number that was already sitting on the person's card. So I made it read the card and the owner's recent notes before writing. I also pulled a time limit that was cutting it off and a rule that let it make one change per message. Ask for several things, including undoing all of it, and it does every part and reports back on each.