Claude Code message prefill sparks training-use speculation
What happened
On October 7, Hacker News reported on Claude Code's feature that prefills the next message after a task finishes, along with Zohaib Ansari's analysis of its potential training use. He speculated that users sending a suggestion directly could form a positive label, that edits before and after the suggestion could form preference pairs, and that predicting a user's next request could help the model learn the order of development work. The author stressed these are only guesses and that he does not know how Anthropic uses the signals; whether sessions are used for training depends on the user's plan and privacy settings. The report does not confirm that feedback from the feature is used for model training.
Written by AI from the coverage · updated 8 minutes ago
Coverage
Follow the reports to see the story from different sides.
- Hacker News front pageClaude Code’s suggested message feature: I think the real customer is the model
Claude Code 会在任务完成后预填下一条消息，Zohaib Ansari 推测用户接受或修改建议的行为可能为模型训练提供反馈。直接发送可作为正向标签，修改前后的文本可构成偏好对，预测用户下一步请求也可能帮助模型学习开发工作的顺序。作者强调这只是猜测，并不了解 Anthropic 如何使用这些信号，会话是否用于训练取决于用户套餐和隐私设置。
Heat over time
Not enough continuous observations to draw a trend yet.