Primary sourceDecoder
View sourceEvent ·Multi-source
Anthropic 提示旧提示词习惯拖慢 Claude Opus 5.5 表现
Anthropic 将 Claude Opus 5.5 设为所有付费计划的默认模型,并同步发布了一份提示词指南,指出许多用户沿用的旧习惯反而在拖慢新模型。指南的核心建议是:放弃"think hard"这类强制深度思考的指令,改从 Medium effort(中等推理力度)起步,让模型自行判断何时投入更多算力。 文章解释,Opus 5.5 的推理控制机制已内置档位调节,旧式的手动加强指令会迫使模型在简单任务上也进行冗长思考,增加延迟和成本。从 Medium 起步、按需升档的做法与模型新的默认行为更匹配。 对开发者而言,这意味着迁移到 Opus 5.5 时需要重写一部分系统提示词,尤其是此前针对旧模型调优的思考链指令。仍在付费计划上使用旧提示模板的团队,可对照 Anthropic 的指南逐条排查。
Current read
如果你在用 Claude,这篇挺实用:Anthropic 官方承认老一套的"think hard"指令对 Opus 5.5 反而有害,建议直接从 Medium 档起步。写提示词的人可以照着指南把旧模板清理一遍,响应速度和 token 开销都会改善。
11 sources57° AI heatLast updated 2026/11/14 16:42:56
Evidence
11 sourcesRelated source 1AI Engineer
View sourceRelated source 2宝玉
View sourceRelated source 3rohanpaul_ai
View sourceRelated source 4kimmonismus
View sourceRelated source 5lmarena.ai
View sourceRelated source 6shao__meng
View sourceRelated source 7IT之家
View sourceRelated source 8claudedevs
View sourceRelated source 9Logan Kilpatrick
View sourceRelated source 10TestingCatalog
View sourceConflict check
Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.