AI Weekly Malaysia

A Hacker News discussion (238 points, 138 comments) links to a blog post arguing Claude Code’s new pre-filled suggested next messages—e.g. “run the tests” or “commit this”—are less a user convenience than a way to collect model feedback. The author says thumbs up/down are sparse and biased, while paid annotators are expensive and lack context; pre-filled suggestions instead predict the next turn from the whole session. Sending a suggestion unchanged is a positive label, and editing it creates a preference pair whose diff shows where the prediction was wrong.

Why: If you use or build AI coding agents, this points to a concrete design pattern: log accepted vs edited next-prompt suggestions as low-friction preference data instead of relying on thumbs up/down. For Claude Code users, editing the pre-filled prompt is not just convenience—it likely produces a higher-value training/eval signal than clicking a feedback button. Builders evaluating agent UX should decide whether to adopt this pattern and whether to disclose it to users.