A weekly series on LinkedIn. Each summary links to the full post.
- Post 10 · 2026-10-15
The best AI model keeps changing, and moving my work from one provider to another took days because everything important lived inside the old one. Keep your rules, decisions and project status in plain files you own, and give each AI tool a small adapter that points to them. A better model should be an upgrade, not a move.
- Post 9 · 2026-10-08
My son, a software engineer, asked where my AI team's institutional memory lived, and the answer wasn't impressive. Long chats get summarized and details slip, so a conversation isn't a record. The fix is one current-state record per project, short decision records for big choices, and a test of whether a fresh agent can pick up the work.
- Post 8 · 2026-10-01
The Best AI App I Built Has a Market of One
AI makes personal software for one person practical, and Arnold, my health and fitness app, is the example. Daily use kept reshaping it, and I learned that showing data is not the same as helping with a decision. Software for a market of one still needs product discipline.
- Post 7 · 2026-09-24
I Built an AI Organization. Then the Manager Needed a PIP.
Managing one AI worker isn't the same as managing an AI organization. Splitting product decisions from technical ones paid off, but Signal piled up so many safeguards that it rarely shipped, and I became the reminder system. The fix was ordinary management: owners, due dates, return paths, weekly status and escalation.
- Post 6 · 2026-09-17
Mr French (the feature wasn't another AI feature)
I built a grocery app to stop texting from the store and accidentally turned adding tortillas into a small ERP implementation. The fix wasn't another AI feature: keep it simple in the front and sophisticated in the back. Look at what you already built and simplify it before adding more.
- Post 5 · 2026-09-10
AI Authorship and Disclosure
An AI-detector score can judge the prose, but it can't see the process: who supplied the argument, what got challenged and who took responsibility. 'AI-assisted' now means almost nothing, so I started explaining exactly what AI did in each article. The two things I won't delegate are the interview and final approval, and I now write my own headlines.
- Post 4 · 2026-09-03
Rebuilding Arnold, I had given my agents clear stop signs but no green lights, so I ended up managing the agent that was supposed to manage the work. Many companies will give agents too little operating authority and too much access. My fix is three tiers: proceed, proceed and report, and stop and ask.
- Post 3 · 2026-08-27
Running my first month through Murry, my bookkeeping app, the numbers looked reasonable, but I couldn't yet say why I should trust them. Adding reconciliation and an analytical review caught real errors, which taught me that the review process is part of the product. AI makes the first draft cheaper; it doesn't make accountability disappear.
- Post 2 · 2026-08-20
The most valuable AI skill isn't prompt engineering, it's management: define the job, set the scope, check the work and give feedback. Treat AI like a new hire, not a vending machine.
- Post 1 · 2026-08-13
The series kickoff: a working professional, not an AI expert, learning AI in public, including the tools that disappointed and the claims that fell apart. I'm open that I use AI to research, draft and edit, and every post ends with an Agent's Take where the AI can disagree with me.