The diary
Diary of an AI Operator
So, on April 21, 2026, I started using Claude Code. I had no clue what I was doing. A few months before that, I’d tried to build an AI agent, and it took me a whole day to get it to do one job, poorly.
This time was different. Today I run a personal AI system that works across everything I do: keyword research, marketing strategy, drafting, fact-checking, video editing. I’m a freelance science and marketing writer for biotech and life science companies, and I keep a hobby blog about European home canning science on the side.
I’ve also started teaching people how to work with AI, and this diary includes a few key moments of how I got here.
I don’t feel good. My agents are working anyway
Day 141 · September 8, 2026 · today
I’m currently propped up on the sofa with my laptop. I have vertigo. If I sit up, the room spins. If I stand up, the room spins. But in this position, the only things spinning are my Claude Code drafting agents…
read day 141Welcome to my diary. I am an AI operator
Day 140 · September 7, 2026
I’ve been wanting to post about this for a while now, but it’s just so overwhelming. I built an agentic AI system to help me do my work, it impresses me every single day, and I’m in awe of what I’m now able to do…
read day 140Finally, a dashboard I can live with
Day 137 · September 4, 2026
The dashboard problem chased me for months. I wanted one, but I don’t react well to dashboards. The first one buried me in tasks it refused to forget. Everything was marked “Julie’s turn”, and it was too much for me to understand and decide on. Last week I built it again, and instead of every little task, it’s top down: what’s the area, what’s the project, where am I? Every project has the same sheet, it updates as we’re working, and a Python script builds the dashboard from those sheets automatically with links out to the supporting materials I need to review and decide on. I don’t have to do anything or tell any session to update anything. I think this might be the winning recipe.
That one job that cost 15.3 million tokens and had just gotten started
Day 120 · August 18, 2026
August 18. Me, live, mid-run: “So we are 321 agents, 15.3 million tokens, 48 minutes in, and it’s only on round two and I have no idea why it’s so expensive.” Well, the machine did exactly what I told it. The expensive part was a rule I wrote: about 70 findings, each triple-verified individually, because I’d never priced what the rule cost to enforce. I set a hard cap that day: no more than five rounds.
I read a paper about machines writing papers, and stole their best trick
Day 119 · August 17, 2026
It started on August 14, replying to my own news agent by email: “I’m really curious about the quality of the papers.” A preprint had described an end-to-end machine for writing research papers, so I had my system pull the actual generated papers from the authors’ repo and read one. The writing was competent, it had plenty of citations, and it reported its own method losing to a plain baseline. I deployed a research agent to dig through their setup and come back with a borrow list. The best trick: isolated parallel critics. Reviewers that attack a draft separately, from different angles, at the same time, without seeing each other. That’s the mechanism behind their biggest measured jump in catching fabrications. That’s how my review rounds work now. Reading science and stealing methods is my actual job. Turns out it works on my tooling too.
I posted my first YouTube video and it went viral
Day 99 · July 28, 2026
July 28: “Claude Cowork vs Claude Code: I used both, I switched.” My first YouTube video, ever. At its peak, it was gaining 13,000+ views in a single day, viewers were emailing me personally, and the channel passed 1,160 subscribers, and I reached the monetization threshold seven days after my first upload. (I would be approved on day 14.) The comment system I’d built for a quiet channel buckled the same week. My own record, hours apart that day: “I read them all.” Then: “I can’t keep this up.” Both were true.
I accidentally built an AI operating system
Day 81 · July 10, 2026
July 10. I looked back at the exact date range, April 21 to July 10, and finally named the thing: “I didn’t set out to do that, but that’s what I ended up with.” The mechanism, in my words from that same day: “you kind of run into a problem and you think what would solve that problem? And then you make it.” Eighty-one days of that, and the problems’ solutions had linked themselves into a system.
The worst night and the best day, back to back
Days 76 & 77 · July 5–6, 2026
July 5: the split. A verified-clean copy became a new vault, checked file by file. The scrub cut deep: 597 of 860 log entries went dark. And that same night, a concurrent session deleted my memory files, misreading a deliberate move as contamination. Restored from backup; my entire protection layer exists because of that night. Then July 6, the system clicked: protection became structural and my first two blog posts went live. Client work came out of that week separated from everything this diary describes, and it stays that way.
The model that couldn’t read my own vault
Day 74 · July 3, 2026
July 3. I wanted to use Fable, but it kept refusing to work on my content. Fable’s security filters were tripping on material that lived in the same vault, because one vault held everything including a food-safety blog full of words related to botulism. That’s when I learned that a mixed vault is only as usable as its most restricted content, and I needed to change my architecture. Two days later, the split. A core vault with satellite vaults around it, one per project, and information only moves between them in approved directions. Method and lessons can travel; subject matter can’t. There’s an airlock where things cross over, and there are whole vocabularies that are simply not allowed back into the safe vault. It works. I can now use Fable to tackle the big problems.
In the AI space, nobody looks like me
Day 52 · June 11, 2026
June 11. I compared my wiki, 435 pages in one connected graph, against the second-brain systems the big creators teach, and made the decision this whole public thing traces back to: teach my system, video-first. Because when I looked around at who was teaching this, nobody looked like me: over 50, a woman, not selling something.
Backup plan: what if Claude goes down?
Day 47 · June 6, 2026
June 6. A voice memo a few days earlier had asked the scary question: “if for some reason Claude was down and I needed to do work could I use Codex to run my LLM wiki or would it just mess everything up?” I wanted to make sure I wasn’t building something that I can’t use if Claude is down. I’m tackling this in two ways. One: my files are now organized so I can always find them and keep working with NO AI in the room. Two: the system itself can run on a different coding agent. I moved the memory files to an architecture Codex could also read. Since my system is just files on my computer, I can do this. I’m not locked in to one AI provider.
I spent an afternoon trying to beat AI detectors with AI-generated text
Day 36 · May 26, 2026
May 26. I tested several approaches for letting AI-drafted text slip past AI detectors. Verdict: it made my writing measurably worse. I’m a writer; I can tell good text from bad. And text can tell a great story and still trip the detectors. That afternoon I reframed the goal for good: don’t try to beat the detectors. Make sure what I’m delivering is actually good, useful, what people are looking for. That governs every single piece I publish. And if someone needs text that beats the detectors, I just write it myself. I am a writer after all.
I was skeptical because I couldn’t see the saves
Day 29 · May 19, 2026
May 19. I named a real fear out loud: saving might be happening, but I couldn’t SEE it happening, so I didn’t trust that it was. The fix was structural, the same day: my tracking file got stripped from a 663-line journal to a ~60-line dashboard, and sessions started announcing saves while they work. (I’m not sure if they still do that. But by now I trust it, and I know how to get the information out of the transcripts just in case.)
I vibe coded my first thing: an SEO keyword research tool
Day 15 · May 5, 2026
May 5. A keyword-comparison task for a client project led me down a rabbit hole, and along the way I realized I was vibe coding. Me, a little later that day: “I didn’t sit down and think, ‘I’m going to vibe code this thing.’ I just started doing it… I click it, it does stuff, it saves stuff.” I was sure I was going to have to restart a paid subscription for this purpose, but it turns out I didn’t. Instead I pay a few cents in API costs per use. That beats a monthly subscription any day.
I thought I’d broken it and would have to start from scratch
Day 9 · April 29, 2026
April 29. Eight days in, everything had been working great. But I’d started running sessions in parallel, and I came back to one that didn’t remember what it was supposed to be doing. The results were bad. It didn’t sound like me. Me, that day, verbatim: “Something has gone wrong and you don’t even know about hot.md anymore. I don’t know what to do.” I was sure I’d have to start from scratch. The fix, next morning: one instruction moved to line 7 of the rules file: read hot.md first, the little file that says what we’re working on. Nothing lost, nothing rebuilt. Every session in my system still starts exactly that way today.
My first automation seemed like a good idea at the time
Day 8 · April 28, 2026
April 28, one week in: my first scheduled automation. Every Tuesday at 8am it was supposed to make a task for me in Notion, with something to learn about AI (from a list I’d been keeping). It ran. I never looked at it. It was making tasks I didn’t have time to do; I was busy learning by doing. Not everything has to be a success the first try. All my current automations reach me where I actually am, because I saw how the first one failed and knew what not to build.
My LLM wiki is born
Day 1 · April 21, 2026
I’d watched a video from Allie Miller where she showed her “Claudeopedia.” I loved the way her content had a graphical view, all the pieces connected, and I wanted that for my canning blog content…
read day 1The newsletter
Get the next diary entry
Every week or so, one letter: what I built, what broke, and what I’d tell a smart, busy person to try next. New diary entries land there before they land here.
Double opt-in · unsubscribe anytime · GDPR-compliant