Back when I wrote technical articles for the Confluent blog, I had this constant argument with the editorial team: They hated that I used the same words to describe the same thing. Especially when it meant saying "low latency" 20 times in a 500 word post. They suggested I mix things up with "performant", "fast", "real time", etc. My opinion is that using different words will be imprecise at best and confusing at worst. Using the same words to describe a concept whenever it comes up, helps readers know that I am still talking about the same thing and not introducing new ideas. For me, this was good writing. Fast forward 6 years, and I'm using LLM that was clearly trained by the same editors who drove me nuts. And I think the harnesses also penalize models for repeating the same words in the output. So every single time I need to ask: "You described this part of the design as 'retry path' and the other part as 'drive forward after failure' - do you mean retry in both cases? or is 'drive forward' somehow different from retry? And no matter how often I say "please use the same terms when describing the same idea", it doesn't help - the LLM remains bound by its training and penalties, and I get to try and decypher what it means each time.