Original Summary

本帖使用社区开源推广，符合推广要求。我申明并遵循社区要求的以下内容： 我的帖子已经打上 开源推广 标签： 是 我的开源项目完整开源，无未开源部分： 是 / 我的开源项目已链接认可 LINUX DO 社区： 是 我帖子内的项目介绍，AI生成、润色内容部分已截图发出： 是 以上选择我承诺是永久有效的，接受社区和佬友监督： 是 以下为项目介绍正文内容，AI生成、润色内容已使用截图方式发出 github.com GitHub - NB-Corp/gpt-with-taste: Better judgment. Better taste. Less ceremony.... Better judgment. Better taste. Less ceremony. System instructions and source-aware steering for GPT in Codex. 原理是系统提示词+hooks，就是系统提示词先大致规定一下，然后每一条用户消息后跟着一些提醒 有关的研究和采用的风格： github.com/NB-Corp/gpt-with-taste docs/research.md master # Research behind the recipe The useful question is not “How do we make GPT sound more confident?” It is “What should the model spend its judgment on?” GPT with Taste redirects that effort toward the user's outcome, supported explanations, and useful work. ## Writing for the reader, not the imagined critic Junchi Liao's Writing for the Reviewer: Defensive Writing in GPT Models compares rewriting and evidence-based writing across GPT and other model families. Its rewriting study uses 77 English computer-science paragraphs. In the controlled prompt comparison, GPT-6 Astra softens 9% of tracked claims under a polishing request and 33% under a request to address likely reviewer concerns. Retraction rises from 0.2% to 9.5%. Repeated review and revision also produces more defensive text. The reader experiment reveals an especially useful mismatch: AI reviewers give the reviewer-oriented Astra rewrites higher overall scores, while four human readers rate them harder to read. The overall AI score rises by 1.22 points on a ten-point scale; ease of reading falls by 0.70 points on a seven-point scale. Comprehension questions receive equally correct answers in both conditions. This changes the editing objective in our prompt: preserve supported meaning and solve the reader's actual question. A review comment should change a claim when it supplies a fact, counterexample, or reasoning gap. An imagined objection earns no extra paragraph. The paper reports GPT-5.6 Sol and GPT-6 Astra at high reasoning effort. Our development cases use GPT-6.1 Sol. We keep those model identities separate when interpreting the results. ## Confidence is not a scoring shortcut EMBER studies uncertainty markers in LLM-based evaluation. Its judges include GPT-4o and earlier GPT models. Adding uncertainty language can change a verdict even when the answer's meaning stays the same. 此文件已被截断。 显示原始文件 有前向测试和调优但是可能不彻底，欢迎佬友调优和pr 1 个帖子 - 1 位参与者 阅读完整话题

  • 情报分类：开源项目与落地
  • 分类依据：内容涉及项目实践、创业、副业或变现
  • 信息来源：服务器 / LINUX DO - 最新话题
  • 发布时间：2026/10/13 11:55:32