V2EX user shares guide on using Claude to optimize GPT custom instructions for less verbose style

On 2026-11-06 10:43, a post appeared on V2EX titled (in Chinese) "Using Claude to optimize GPT custom instructions to improve verbose writing style." The post is a share by a community member. It claims that having Claude write custom prompts (that is, custom instructions) can effectively fix the low information density, formulaic phrasing, and overly confident tone found in GPT's answers.

The poster then analyzes why the model's prose tends to drag on, offering three explanations. First, training rewards favor answers that "look rigorous": the human raters or the reward model tend to score highly responses that are comprehensive, non-absolute, and consider every case. Constructions like "is not the same as," "does not mean," and "cannot be based solely on" appear thorough and non-dogmatic by nature, so they get reinforced. Second, the model over-defends against misinterpretation: it has been trained to fear twisting what the user means, to fear saying too much, and to fear being faulted, so it develops a habit of knocking down a straw-man viewpoint no one raised before giving its conclusion. For example, the user never said "popular means simplistic," yet the model insists on denying it first — this is the source of all the "does not mean" phrasing in a passage. Third, "not A, but B" is inherently a high-payoff construction; contrast structures read as insightful, and since they are rampant in Chinese social media and WeChat-style columns, the model has learned a great deal of them. Generated text is then fed back to train newer models, so the style compounds. Self-reinforcement happens within the context as well: once the pattern appears in the first few sentences, later continuation becomes more likely, so the writing becomes ever more extreme, losing its logic and keeping only the sentence pattern — the poster's example, "I cannot prove you need to uninstall me, nor can I prove you do not," is pure syntactic inertia.

The post concludes that this style ismocked across the web mainly because it offers very low information density while sounding highly assertive. The poster also notes that these are only guesses — he does not know how OpenAI trains internally. As of now, the topic has only this single post and no reported follow-up.