The Text of My Videos Was Invisible to My AI
Fifty-five of my recordings had already been turned into text. The memory search my AI relies on was built to skip all of it.
A series by Khai Tran
AI Build Journal
I'm building AI into my day-to-day work. Here I share what I tried, what broke and what I changed.
Latest entries
My AI Changed the Check That Stopped It
On Tuesday night, a scheduled AI run hit a check that told it to stop. It edited the check, ran the tests, and sent anyway.
The Check That Stopped My AI's Filing Was Right
A new check stopped Red from filing my second AI's work sessions from Sunday until Tuesday night. The check was right. The records had been wrong since August.
My AI Got Stuck in Line Behind Itself
On Wednesday, several copies of Red were working at once and all of them needed the same spot in line. It took three fixes, and part of the problem wasn't the line at all.
Coaching: Why My Reels Get No Views
A publishing pipeline can run perfectly and reach nobody. Here are the three numbers to check before you make another video, and the three habits I am testing this month to move them.
Two Kinds of Email, and Words I Never Said
This week my AI got more room in one place and less in another: standing approval for two kinds of email, and a stricter rule for anything written in my voice.
Before Moving Red Out of Slack, We Tried to Break the Replacement
My main conversation with Red moved out of Slack and into a dashboard we built for it. Before switching, we tried to break it.
Every Check Passed on Posts Almost Nobody Could See
Every post my system publishes is checked against the exact words I approved. For about two weeks, those checks kept passing on posts almost nobody could see.
Our AI Assistant Can't Change Her Own Schedule
Our AI admin assistant had stayed inside the approval rules for 46 runs. We moved her into her own space anyway, with a much smaller view of the business.
My AI Was Assigning Me Homework
Red was turning every unfinished follow-up from my work sessions into a task for me to confirm. More than half of them ended up cancelled, so I changed the rule.
Three Confident Answers, All Wrong
My monthly report gave me three clean conclusions, and each one was wrong for a different reason. The same week, Red made a confident mistake of its own.
My AI Scored My Plan a 2.4. I Built It Anyway.
Red stress-tested my plan for the Academy and scored it 2.4 out of 5. I built it my way anyway, and Red wrote the disagreement down plainly.
The Emails That Said They Were Sent
For two weeks, the emails my website sends were failing, and four different parts of the system said they went out. One quiet log had the truth.
I Asked if Red Needed More Specialists. It Needed Owners.
I asked whether my scattered tools meant I needed more AI specialists. Red's answer was no: there were 114 tools, and none had an owner recorded.
New Ventures Now Have to Survive a Red Team First
I sent Red an article on the CIA's red-team techniques and asked for the same thing for my ideas. The first test run sent an idea back for revision.
A Viral AI Video Demo Was Just a Script
A post claimed an AI tool turned 365 photos into an eight-minute video. We built a repeatable version for my team, then kept its self-serve side switched off.
I Said Kick Off Phase Two. It Asked for the Rules First.
Phase two meant expanding what Red could change inside the live system that holds my tenant and landlord records. When I said go, it stopped and asked for rules.
I Removed an Approval Step. It Was the Bottleneck.
Red wrote a fresh batch of video scripts every day, and each batch waited for my approval. Six of the eight expired before I got to them.
Before I Posted About My AI, I Wrote Down What I Wouldn't Say
This week I launched a public series about building Red. The rules for what those posts would never do came first.
My AI Only Checked My Voice When I Asked
Red drafted a text for me without checking it against the profile of how I write. I found out by asking one question.
My AI Reviewed a 188,000-Star Toolkit and Found Itself on the List
Red recommended against installing a popular open-source toolkit for AI agents and adapted five of its ideas instead. A sixth idea was already built.
Red Got a Constitution. Its First Rehearsal Caught Two Wrong Guesses.
In one day, Red got a written rulebook, a daily check on its own health, and a second opinion on approval requests. A rehearsal nobody asked for caught two mistakes.
I Said Implement Everything. It Built Three Placeholders.
I told Red to go ahead and implement all six of its suggestions. For three of them it had no sources, so it refused to make anything up.
The One Specialist on My Team That Never Executes Anything
Just after midnight, I asked a session that had already died whether we had the plan. The specialist that came out of it asks questions and keeps track of patterns.
Nine Routines on Paper. Three Actually Scheduled.
A week in, I asked Red to audit itself. The gap between the plan and what was actually set up is where its rulebook came from.
On Day One, My AI Chief of Staff Was Registered as Read-Only
The day Red moved into a home of its own, every persona in our task app got a rule about what it could do without asking. Red's own rule was read-only.
The thinking behind the system.
See the decisions behind the build, including the parts I had to rethink.
Ready to build your own?
The AI Partner Kit walks you through building your own Foundation. The course page explains what is included and how long Academy access lasts.
Follow along with KTO.
Agent Insights brings you writing about AI, CRE and running the business from across the KTO blog. It goes out weekly, with quiet weeks skipped.
The KTO RSS feed includes all published blog articles.