An update on recent Claude Code quality reports
Summary (EN)
Anthropic published a postmortem-style engineering note on April 24 addressing user reports that Claude’s quality had worsened in recent weeks. The company said its investigation traced the issue to three distinct product-layer changes affecting Claude Code, the Claude Agent SDK, and Claude Cowork, while clarifying that the API and inference layer were not impacted. Anthropic said the first issue came from lowering Claude Code’s default reasoning effort from high to medium on March 4 in an attempt to reduce latency. The second came from a March 26 change intended to clear older thinking from long-idle sessions, but a bug caused the reset to recur every turn, making the system seem repetitive and forgetful. The third was an April 16 system-prompt change meant to reduce verbosity, which in combination with other prompt edits hurt coding quality. Anthropic said all three issues had been resolved by April 20 in version 2.1.116 and that subscriber usage limits were being reset as of April 23. The note is significant because it offers unusually specific operational transparency from a major model provider about how product defaults, memory handling, and prompt-layer tuning can materially affect perceived model quality even when the underlying inference stack remains unchanged. It is an important signal about the fragility of AI product experience at the interaction layer.
Summary (ZH)
Anthropic 于 4 月 24 日发布了一篇带有事后复盘性质的工程说明，回应近期部分用户关于 Claude 表现变差的反馈。公司表示，调查结果显示问题并非来自底层 API 或推理层，而是来自产品层的三项独立改动，这些改动影响了 Claude Code、Claude Agent SDK 和 Claude Cowork。第一项问题源于 3 月 4 日将 Claude Code 的默认推理强度从 high 调低到 medium，目的是缓解部分用户遇到的高延迟问题，但这被证明是错误的权衡。第二项问题来自 3 月 26 日的一次会话优化，本意是清理长时间闲置会话中的旧思考内容以降低恢复延迟，但一个 bug 导致这一清理动作在后续每一轮都重复发生，使模型显得健忘且重复。第三项问题则是 4 月 16 日为降低冗长度而加入的一条 system prompt 指令，该指令与其他提示词调整叠加后损害了编码质量。Anthropic 表示，这三项问题已在 4 月 20 日的 2.1.116 版本中修复，并自 4 月 23 日起为所有订阅用户重置使用额度。该说明的新闻价值在于，Anthropic 罕见地公开展示了产品默认参数、记忆处理方式和提示层调整如何在不改变底层模型的情况下显著影响用户感知质量，是理解 AI 产品脆弱性的关键案例。
Source