Been thinking a lot around this lately. Your engineers can prototype anything in hours. Your domain experts still need weeks to make it work well. Coding got 10x faster. AI improvement cycles didn't. Guess which one can determine who wins?

Casually deploying a fix to prod for my fun project while on a flight so I can listen to podcasts created from my newsletters again. 😅

The latest Acquired podcast intro music is fire.

My Spotify is entirely podcasts, thunderstorm sounds and F1 soundtrack now.

Recent lesson: Have a plan increases fragility. Always be planning.

Huh

I was curious to try using gpt-5 high reasoning to run a commit and push command. A ton of thinking went into that. 😅

For those building agents, how reliable are your agent in production?

I wish Mac OS can adjust between dark and light mode based on surroundings, in addition to just time of day. When using laptop on a flight with a dimmed cabin, it causes a ton of eye strain. Yes, I am that lazy to go to system settings. It should just work, right?

Finally time to move the lengthy instructions into skills...

Codex 5.1 is so good that I am getting rate limited lol.

Claude Cowork is amazing, but I think the challenge remains in context management, and now for non-technical folks. I have been using Claude in the "cowork" way for more than a year, and I still forget to remove obsolete files. Now imagine an average Mac OS desktop or folder.

Claude to brainstorm and write specs, Claude code to write, Codex to review and fix bugs. Works like a charm lately.

AI writes all the code. We still measure in lines changed. Why? Because humans need to feel scope. But is it the right unit anymore? Show me features impacted, not files touched. Who's building the semantic layer on top of the codebase?