How to Break an LLM's Bad Verbal Habits

Logosofia no Dia a Dia · May 31, 2026 · 34m

Every large language model develops a fingerprint of overused phrases — "delve," "failure modes," "it's worth noting." Blacklists leak into output, regex silently inverts meaning, and feeding previous episodes into context only compounds the problem. This episode explores three practical approaches to breaking the reinforcement loop: compressed style summaries with decay, dual-model editing architectures, and context-aware regex using dependency parsing. No silver bullets, just real trade-offs for production pipelines. Episode 3171 — myweirdprompts.com/n/3171 (permanent short link…