Keep an EDM festival anthem exactly as it is and change one word, the name shouted over the drop, to a friend's. Pure AI, no recording. In June 2026, with models that write whole songs and clone voices on demand, this should be a one-evening job. It was not.
The research told me who to interview and ranked them. Then I had to actually run the calls. A 30-minute screen usually wastes both people's time, so I built the same thing one stage later: a per-candidate interview page my agents fill from the research, with every question pre-tagged to the score it feeds.
I swapped the username scanner in my candidate-research workflow for one that searches seven times as many sites. The reach wasn't the upgrade. The upgrade was that finding an account was never the bottleneck. Telling whose account it is, was.
I built a parallel-agent workflow to make candidate background research faster. Sixty-four candidates later, the speedup is the least of it: running the investigations side by side turned a noisy classifier into a confident one, and then into a way to rank the candidates who are real.
For fifteen years I've built a custom Chrome theme at every company I've worked at, and always built it badly, because it was never worth doing well. This time AI made it the best version I've shipped, for the same one evening of work. The theme is trivial. The category of task it unlocks isn't.
I stopped giving my AI assistants rules and started giving them reading material. The shift is from instruction to artifact. The em-dash regex is the catchy part. Loading files as books-the-system-reads is the durable one.
The muscle that ships a company from zero is not the muscle that keeps a company aimed. I have shifted between modes more than once. Each time, the habits that worked before were the habits I had to break first.