Seven weeks into running everything from my phone, the AI isn’t what slows me down anymore.
It’s me.
When I started this experiment — managing Transkrybe, content across seven platforms, a job search pipeline, and a book promotion all from my iPhone using AI agents — I figured the hard part would be the technology. Getting the agents to actually do things. Getting the tools to talk to each other.
That stuff got figured out faster than I expected. The harder thing — the thing nobody warned me about — was teaching the AI to sound like me. To know what I care about. To understand that when I say “draft a response to this recruiter,” I don’t mean a LinkedIn influencer hitting engagement goals. I mean something that sounds like a human wrote it at 6am while drinking bad coffee.
That took weeks to get right. And the way I got there is a little embarrassing.
The Amnesia Problem
Everything I run uses Claude — Cowork sessions on my phone, scheduled tasks, Slack agent workflows. And by default, Claude doesn’t remember anything. Every session starts fresh. You explain yourself from scratch. Same context, same preferences, same project details — every single time.
It’s like having an assistant who gets amnesia overnight. Every morning: “Hi, I’m Claude.” Right. I know who you are. Let me spend the next ten minutes re-explaining what we’re building.
The fix I landed on is a file called CLAUDE.md. A permanent briefing document that loads at the start of every session. Credentials, project context, how I like things formatted, what NOT to do. Over the last seven weeks, that file has grown from about 200 words to over 1,000.
And here’s the thing — almost every addition came from pain. Claude did something wrong. I thought “I need to make sure that never happens again.” I wrote it down.
What’s Actually in There
Some specific examples, because I think the actual details matter more than the concept.
My blog voice. “No ‘it’s worth noting.’ No parallel bullet structures. No em dashes for emphasis. Start sentences with And or But sometimes. Use fragments.” I added that after getting back three drafts in a row that read like a McKinsey deck.
Browser rules. “Always Brave, never Chrome. Close tabs when you’re done. If tab count exceeds five, close unused ones first.” That went in after a session froze because there were 47 open tabs in Brave. 47. The browser locked up, automation stopped, I lost twenty minutes to recovery.
Content approval. “Never post without approval. Always draft first.” Week one. After a miscommunicated instruction nearly auto-published a half-finished LinkedIn post. That would’ve been bad.
Outreach voice. “Jesse’s been in tech for 20+ years and has opinions. Write like that. Not formal, not precious.” Added after early recruiter outreach messages were getting sub-10% response rates. After the update? Closer to 25%.
None of this is sophisticated. It’s just documentation. But the cumulative effect of seven weeks of “write that down” is that sessions now start cold and immediately run warm. The re-orientation tax is almost gone.
What It Changed for Content
The memory system matters most for this blog — and for the LinkedIn posts that come from it.
My weekly posts start as a brief now. Project updates, what happened this week, what angle to take. Claude drafts from that brief with full context on my voice, the series history, what topics I’ve already covered. The first draft is usually 70% there. Eight weeks ago? Maybe 30%.
Same with LinkedIn. Three variants go out per week — a Monday hook post, a Wednesday contrarian take, a Friday numbered list. Each one actually sounds like me now, because the context file includes specific phrases I use, tones I avoid, and examples from past posts that landed well. The agents don’t have to guess. They have a model.
The job search pipeline got smarter too. My outreach agent sends first-contact messages to recruiters, and when I added voice guidance to the memory file, response rates went up measurably. Small tweak. Real result.
The Parts That Still Break
I’d be lying if I said it runs clean.
Biggest ongoing issue: context window limits. CLAUDE.md can’t be infinite, and some sessions hit length limits before I get through everything I need. I’m still figuring out how to chunk this — probably breaking the memory file into modular pieces that load selectively based on task type. Haven’t solved that yet.
There’s also what I call the “I thought you knew that” problem. I’ll get back a response from an agent that misses something obvious — only to realize I never actually wrote it down. I just assumed it was implied. The more you rely on the system, the more you find those gaps.
And Transkrybe is its own context headache. The product’s codebase lives in Modal. The memory file knows the project exists, but doesn’t have deep technical context on the actual code. So when I’m debugging with Claude, I still have to paste relevant files. That’s the next thing I want to solve — some kind of persistent code-layer context that loads with the project. No clean answer yet.
What “AI CEO” Actually Means
I keep using that phrase, and I’ve been thinking about what it really means in practice.
It’s not about having AI make decisions. That’s not the point. It’s about offloading the execution layer so you can stay in the decisions layer longer. Less time re-explaining context. More time actually thinking.
The memory system is what makes that work at scale. Without it, you’re re-briefing constantly. Losing ten minutes here, twenty minutes there. The agents are capable, but they can’t help you if they don’t know who you are or what you’re trying to build.
With it, you get something closer to real delegation. The kind where you hand something off and trust it to come back reasonably good — not perfect, but good enough to edit.
Seven weeks to get here. Still building.
If you’re running your own version of this — agents, memory files, whatever system you’ve landed on — I’d genuinely like to know what you’re doing. Drop it in the comments.
