OpenAI 官方给 Astra 列了一份禁用词清单

MIKE @mikenevermiss原推发布于 2026.09.15

在 X 上打开原推

OpenAI wrote a list of words its own AI is told to avoid. anyone who reads AI writing knows these words "Delve." "Foster." "Leverage." "It's worth noting." they show up so much that readers can notice them in seconds and start doubting everything else they read. OpenAI noticed it too. the official guide for GPT-6 Astra, their newest model, includes a writing instruction that lists these words and tells the model to stop using them. the guide also explains why. OpenAI says Astra can fill its answers with bullet points, tables and too much formatting, while repeating the same phrases from one chat to another. the instruction also covers writing habits, not just single words. it tells Astra to avoid the dramatic style where a sentence says something "isn't about X, it's about Y." It removes ending lines that start with "In short" or "Bottom line." It also bans made-up labels that sound technical but don't really mean anything, and tells the model to explain what it's doing using simple words. OpenAI also gives a second instruction for more natural writing. Astra should use short paragraphs with one main idea, use familiar words and real examples, put the main point near the top, and only use lists when the information actually needs a list. these rules matter more with this model. OpenAI says Astra follows instructions better than any model they have released before, so a style rule you set once can stick throughout the whole conversation. If you use Astra for posts, emails, reports or client work, copy both instructions from the guide and add them to your custom instructions or system prompt. your answers will start closer to something a person would actually write, and you'll spend less time cleaning them up. OpenAI admitted the problem and published the fix in the same document. my full guide to running Astra at max output is right below👇

这条是引用着另一条写的：x.com/mikenevermiss/status/2098270564016091458