Anthropic 事後檢討：三個改動讓 Claude Code 變笨，官方怎麼修
An update on recent Claude Code quality reports
摘要分析
2026-04-23 Anthropic 公開事後檢討，回應 Claude Code 品質下滑抱怨。查出三個改動：3/4 預設 reasoning effort 從 high 降到 medium；3/26 清除舊 thinking 的改動有 bug 變成每輪都清，導致健忘、重複、更快耗盡用量；4/16 系統提示要求回答精簡反而降低寫程式品質。三者均已回復，API 未受影響。改善措施含每次系統提示改動做逐模型評測與分階段上線，並重置用量上限。啟示：模型「變笨」常是 harness 設定問題。
- 品質下滑來自三個 harness 改動，不是模型本身；API 不受影響
- reasoning effort 預設值會大幅影響結果，可用 /model 或設定調整
- 清除 thinking 歷史的 bug 讓 Claude 健忘且更耗用量
- 要求「回答精簡」的系統提示會降低寫程式品質
- 官方新增逐模型評測、消融與分階段上線
On 2026-04-23 Anthropic published a postmortem answering reports that Claude Code had gotten worse. Three unrelated changes were found. On March 4 the default reasoning effort moved from high to medium to cut waits; users preferred intelligence, so it was reverted April 7. On March 26 a caching optimization meant to clear old thinking from long-idle sessions had a bug that cleared it on every turn, making Claude forgetful and repetitive and draining usage limits faster; fixed April 10. On April 16 a system-prompt instruction to shorten responses, combined with other prompt changes, reduced coding quality; reverted April 20. The API was unaffected. Safeguards now include more staff on the public build, per-model evals and ablations for every system-prompt change, soak periods and gradual rollouts, plus a usage-limit reset for all subscribers. The lesson for users: 'dumber' behavior is often a harness setting, not the model.
- Three harness changes, not the model; API unaffected
- Default reasoning effort strongly affects quality
- Clearing thinking history made Claude forgetful
- Terse-response prompts hurt coding quality
- New per-model evals, ablations and gradual rollouts
關鍵字：postmortemClaude Codereasoning effort系統提示品質
來源原文：https://www.anthropic.com/engineering/april-23-postmortem
本頁為報導內容的摘要與分析，版權屬原媒體所有；請以原文為準。