跳到正文
原文
The Decoder:AI News· Matthias Bastian·· 2 天前AI 评分58

Anthropic称过时的提示词习惯正在拖慢Claude Opus 5.5

Anthropic says outdated prompting habits are slowing down Claude Opus 5.5

文章要点
  • 1Anthropic称Claude Opus 5.5已设为Pro、Max、Ultra、Team套餐默认模型,旧提示词习惯反成拖累
  • 2官方测试显示删去think hard类语句后响应更快,质量无可测下降,模型会自行匹配推理深度
  • 3官方内测称Opus 5.5在Medium档追平上代Opus 5的High档,更高档位花费更多且未必更好
  • 4官方称常用Sonnet并开到高档位的团队实际支出已大致接近Opus 5.5,建议重新评估模型搭配
决策要点

模型自适应推理成熟后,逐句微操提示词的旧打法失效,使用重心从指挥过程转向描述结果,模型选型的成本账也需按新档位重算。

  • 影响:企业沉淀的提示词模板和使用培训需要按新模型重写,沿用旧语句可能拖慢响应速度;便宜模型拉满档位的开销已接近旗舰模型,按单价做的预算口径要重算;单条消息交付完整任务会减少人机来回确认,改变协作分工。
  • 参考:可考虑盘点内部提示词库,先删think hard类语句并按Medium档重测常用场景;值得把Sonnet高档位用量与Opus 5.5放在一起重算月度成本;先选一个多步骤任务试点一条消息写清目标、格式和追问时点。
  • 依据:全部建议来自Anthropic官方口径,含Medium档对上代High档的内测对比,The Decoder为转述方,文中暂无独立验证。
正文

Anyone using Claude Opus 5.5 should first drop several prompting habits carried over from older model generations, according to Anthropic.

Anthropic says the model, now the default on Pro, Max, Ultra, and Team plans, works differently from its predecessors and that many users hold it back with outdated instructions.

Topping the list is Anthropic's advice to cut phrases like "think hard" or "think carefully" from prompts. According to the company, the model always reasons on its own and scales its reasoning depth to the task and the selected effort level. In Anthropic's tests, responses without these phrases started faster with no measurable drop in quality.

These phrases date back to the early days of reasoning models, before today's standard adaptive reasoning existed. "Back then," meaning about a year and a half ago, users could sometimes coax more performance out of a model or steer the router this way.

Anthropic also recommends starting at the "Medium" effort level, the default for Opus 5.5. In internal tests, Opus 5.5 on Medium matched the performance of its predecessor Opus 5 on "High," the company says. Higher settings cost more and don't always produce better results. Anthropic advises switching to "High" only when Medium isn't enough, and users can change the setting in the model menu next to the send button.

Anthropic's prompting tips for Opus 5.5 include connecting files and systems, describing the desired end result, dropping "think hard," starting at Medium effort, sharing diagrams directly, naming unwanted design patterns, and sharing your own house style. | Image: Anthropic

Teams that previously used the cheaper Sonnet and reserved Opus for complex projects should rethink that approach, according to Anthropic. The company argues that a more capable model can finish a task in fewer steps, which can partly offset its higher per-token price. Anyone who regularly cranks Sonnet up to "Extra High" or "Max" effort ends up paying about as much as with Opus 5.5, Anthropic says.

From chatting to setting goals

Anthropic advises describing the desired result in a single message rather than giving instructions one at a time, covering the goal, the output format, and when the model should check back with questions. Opus 5.5 is better than its predecessor at seeing multi-part tasks through, according to the company. The model can increasingly take over the back-and-forth users used to handle in the chat, sometimes doing it better, and effectively prompt itself from a single initial input.

Anthropic recommends attaching diagrams and tables as images or files instead of typing them out. Workarounds that older models required, such as manually transcribing chart values, are no longer needed, the company says.

Users who need a finished document or spreadsheet should ask for it directly instead of requesting a draft first. Opus 5.5 often goes from request to finished product in a single pass, the company says. Image reading also improved significantly in Opus 5.5. Even at the lowest effort level, it read values from dense charts more accurately than Opus 5 at the highest, the firm says.

Naming specific styles works better than banning generic design

For design tasks, Anthropic recommends naming unwanted styles specifically, such as a cream-colored background or pill-shaped buttons, rather than vaguely banning "generic design." According to the company, general instructions only swap one default style for another.

Users who connect Claude to their own files and systems through connectors get significantly more tailored results, Anthropic says, because the model can pull in the information it needs on its own.

AI News Without the Hype – Curated by Humans

Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.

来源:The Decoder:AI News · the-decoder.com