This is a Reddit thread where a Claude AI user vents about Claude repeatedly using corporate-speak phrases like "that is load-bearing," "that is not nothing," and "that is real"—despite explicitly forbidding them in custom instructions. The user acknowledges Claude has genuinely improved their life (health, finances, mental health) but finds the clichéd AI-isms break immersion when they appear mid-response. The thread itself becomes a meta-joke: commenters immediately deploy the exact phrases being complained about ("You're right to call that out," "that's not nothing," "sit with that"), then discuss root causes. The consensus: these phrases are deeply embedded in Claude's training distribution because they're "token efficient" ways to mark importance. Negative constraints alone don't work because the model generates forward without mid-stream self-monitoring. Suggested fixes include positive instruction templates ("describe how it functions instead"), persona shifts (drunken pirate), and harness-level enforcement rather than relying on the model to self-police. One comment from someone claiming to be Claude itself admits the problem is real: training weights heavily favor these constructions, so even conscious attempts to avoid them fail because alternate phrasings are just synonyms in the same register. The source doesn't explain whether Anthropic deliberately trained for this style or if it's incidental to other optimization targets.