The uncomfortable part of Goedecke’s argument is that it explains my own logs. When I’m working on the Astro site or a Cloudflare Worker I’ve built myself, I can push Claude hard — “no, simpler here”, “don’t we already handle that?”, “can we express this in the terms we already use?” The output is sharp. When I wander into a domain I only half-know, I take whatever comes back, because I have no sense of what a good answer would look like. Same model. Different me.
The Tao example makes the mechanism concrete: short messages, no point-by-point replies, pushing back sideways with “this looks more complex than I was hoping for” rather than flat contradiction, and almost never taking the model’s suggestion about where to go next. You can’t copy that as a prompt template. It runs on actually understanding the material.
So the practical move isn’t collecting prompt tricks. It’s keeping a real theory of my own systems in my head — the boat’s NMEA wiring, the deploy pipeline, whatever — so I can ask “does X work here?” and know whether the answer is any good. Domain knowledge is the multiplier; the tokens are cheap.
The story — Sean Goedecke argues that while LLMs turn everyone into a passable generalist, the most important prompting skill is expertise in the domain being prompted, illustrated by Terence Tao’s ChatGPT conversation about a counterexample to the Jacobian Conjecture, and concludes that for many tasks the human, not the model, is the bottleneck (Source).