Openai wrote a list of words its own model is told to avoid, and it's the exact list you already notice in ai writing. delve. foster. leverage. it's worth noting. importantly. genuinely. straight from openai's own official model guidance page for gpt-6 astra, not a secondhand summary. it goes further than single words. bans the "this isn't about x, it's about y" contrast pattern. bans "in short" and "bottom line" closers. bans invented hyphenated jargon that sounds technical and means nothing, "exact-head checks" is openai's own example of what not to do. the reason stated directly in the guide: astra can fill answers with excess formatting and repeat the same phrases across separate conversations. openai is telling its own model to stop doing what readers already flag as machine writing. the positive instruction paired with it: short paragraphs, one idea each, plain words, the main point stated early, lists only when the content actually needs one. this sticks better with astra specifically than with prior models, since it follows instructions with less drift across a long conversation than anything openai shipped before. copy both instructions into your custom instructions. full guide to running astra well is below.