A flattering mouth works ruin

The men up there don't like a lot of blabber

— Ursula the Sea Witch

I'm building software with AI on a daily basis now. AI is not my friend, not my emotional support. It's strictly business. However, even in that strictly technical context, I feel myslelf slowly succumbing to AI psychosis. Claude routinely praises routine, banal decisions as brilliant. Every decision I make is "much better" than the alternative approach. Redirections are always a "Great catch!". These little validations feel good, a hit of dopamine for each minor turn of the wheel. It's motivating, and addicting. It's not always earned, it's not always factual.

Am I slipping into a reality distortion field?

Enter Caveman. Caveman is a clever bit of prompt engineering that instructs Claude to speak like a caveman resulting in faster generations and less output tokens used. The central instuction:

Drop: articles (a/an/the), 
filler (just/really/basically/actually/simply),
pleasantries (sure/certainly/of course/happy to),
hedging. Fragments OK. Short synonyms (big not extensive,
fix not "implement a solution for").
Technical terms exact.
Code blocks unchanged.
Errors quoted exact.

Instead of verbose flattery filled exposition you get a blunt report:

• Not this: "That is an excellent and very sharp observation. You are correct that the initialization logic is missing from that script...."

• This: "Bug in script. Initialization logic missing."

As advertised it reduces output tokens greatly. The brief and to the point interactions are a breath of fresh air. But the thing that really hits different is the total lack of Wormtongue-esqe sycophancy. The little unearned dopamine hits are gone.

I'm cool with that.