'Claude 回顾':你的AI聊天历史记录就像Spotify年度总结

qimuai 发布于 阅读:37 一手编译

'Claude 回顾':你的AI聊天历史记录就像Spotify年度总结

内容来源:https://lifehacker.com/tech/claude-reflect-ai-history?utm_medium=RSS

内容总结:

AI公司推出“反思工具”:像音乐年终总结一样分析你的AI使用习惯

近日,人工智能公司Anthropic为其AI助手Claude推出了一项不同寻常的更新——名为“反思”(Reflect)的新工具。该功能旨在帮助用户审视自身与AI的交互模式,回答诸如“AI使用频率是否合理”“何时该交由人类处理”等问题。

据了解,Reflect的工作原理类似于音乐应用Spotify的年度总结(Spotify Wrapped),可汇总用户过去一个月、三个月、六个月或一年内的活动记录。经实测,该工具生成的总结基本客观且具有洞察力。目前,该功能已在Claude的网页版和桌面应用中上线,面向免费、Pro及Max用户开放,但尚处于测试阶段。

使用Reflect前,用户需先开启“记忆”(Memory)功能。在网页端点击个人头像进入“设置-功能”,或在移动端通过菜单进入“功能”选项,确保“搜索和参考聊天记录”开关已启用。随后,在网页或桌面应用的个人头像设置中选择“反思”,系统将自动生成最近一个月的使用总结,用户可通过右上角下拉菜单调整时间范围。

测试者发现,该工具能准确识别自身的主要使用场景——如开发有助于工作的小型网页应用,以及测试Claude的各类功能(多数情况仍与工作相关)。不过,总结会忽略最近对话,仅统计完整月份的数据,且倾向于提取长期趋势而非单次简短聊天。

向下滚动,用户可看到更详细的分析:Claude会将使用时间按类别拆分,并评估用户的“AI流畅度”——即提示词(prompt)的编写能力。例如,在“任务委托”和“判断AI输出准确性”方面得分较高,但“描述能力”有待提升。系统还会以具体案例指出,何时可以优化提示词以获取更精准的回答。

尽管测试者认为这些总结不太可能改变自身使用习惯,但将其视为一种有益的“自我审计”方式——正如Spotify年终总结能提醒你听了太多90年代流行歌曲,Reflect也能帮助用户重新审视AI使用的边界,尤其适合重度用户。不过,测试者也指出,AI生成的总结有时过于追求叙事流畅性,可能会牺牲部分准确性。

值得注意的是,若用户发现AI使用过度,Reflect总结中附有“设置安静时间和休息时间”链接,可跳转至应用设置中的“时间与专注”板块,在此可限制屏幕使用时长,并设置定时提醒,鼓励适度使用AI。

中文翻译:

我们现在早已习惯了人工智能公司不断推出模型升级和新功能,但克劳德最近的一次更新带来了一些更不寻常的东西:反思工具,它可以回顾你使用这款人工智能机器人的方式。

联合创始人表示,克劳德反思功能的推出是为了回答以下问题:人们应该多久使用一次人工智能?如何才能最有效地使用它?什么情况下人工智能适合处理任务,什么情况下最好交给人类?此外,也许还包括:你是否让人工智能替你做了太多思考?

实际上,它的运作方式有点像年度音乐报告,汇总你在该应用过去一个月、三个月、六个月或一年内的活动。根据我的测试,它的总结大多公平且有见地。

反思功能现已面向免费版、专业版和最大版用户在克劳德应用中推出,不过联合创始人目前将其标记为测试版功能。

开始使用克劳德反思

在使用克劳德反思功能之前,你需要先启用记忆功能:在网页端,点击你的个人资料图标(左下角),然后选择设置 > 功能。在移动应用中,点击菜单按钮(左上角),然后点击个人资料图标,再选择功能。如果搜索和参考聊天开关已开启,你就可以开始使用了。

当你准备进行反思时,你需要在网页或桌面应用中操作。点击你的个人资料图标(左下角),然后选择设置,再选择反思。系统默认会自动生成过去一个月的使用摘要,但你可以通过右上角的下拉菜单更改时间段(点击圆形刷新箭头图标以加载新的摘要)。

我必须事先承认,我倾向于定期删除我的克劳德聊天记录——出于各种隐私和安全原因,我不喜欢留下长长的数字足迹——所以我无法告诉你克劳德翻阅数月对话会是什么样子。或许挖掘更大的数据集会带来不同的结果。

我可以告诉你的是,克劳德对我过去一个月人工智能使用的回顾仍然很有趣:它准确地指出我花了很多时间开发有助于工作的小型网络应用和小工具,当然还有(大多数时候也是为了工作目的)测试克劳德的功能。

然而,它确实会忽略你最近的对话,具体取决于日历:如果你处于七月中旬,你的“过去一个月”摘要只会截止到六月底,因此摘要中只统计完整的月份。它似乎也会忽略较短的、一次性对话,而更倾向于关注较大的趋势。

深入挖掘数据

你的反思以摘要和显示每日活动的图表开始,但如果你向下滚动,还有更多内容可供探索。克劳德会将你的时间分解为你花费时间的类别(比如在我的例子中是 Chrome 的新标签页扩展),并评估你的人工智能熟练程度——基本上就是看你提问的水平有多高。

我在向聊天机器人“委派任务”方面的得分很高,在评估人工智能输出方面的“判断力”也很高,但显然我的“描述”技能还有提升空间。克劳德指出了一次我本可以更精确地提出指令的情况,并给了我一个如何优化指令的例子。

我不确定我在克劳德反思中读到的任何内容会改变我使用人工智能的方式,但我认为偶尔用它进行自我审视是值得的。就像年度音乐报告可以揭示你听了太多 90 年代流行音乐一样,反思功能可以帮助重新评估你使用人工智能的界限——尤其是对于重度用户。

我确实发现某些总结有点过于美化:像克劳德这样的人工智能机器人倾向于写出流畅连贯、可以完美收尾的叙述,而不是更准确但可能更尴尬的报告,这在这一点上再次显现。不过,总体而言,它似乎给了我公平的评价。

如果你确实觉得你使用克劳德的行为有些失控,反思摘要中包含一个设置静默时间和休息的链接,该链接会引导你进入应用设置中的时间和专注部分:在这里,你可以为自己设定使用人工智能的屏幕时间限制,并让克劳德时不时地提醒你休息一下,暂缓提问。

英文来源:

We're now well used to AI companies pushing out model upgrades and new features, but one of the most recent updates to Claude added something more unusual: a Reflect tool that looks back on how you've been using the AI bot.
Anthropic says Claude Reflect has been introduced to answer questions like, how often should someone use AI? How can it be used most effectively? When is AI suited to a task, and when is it better left to a human? And also, perhaps, are you letting AI do too much of your thinking for you?
In practice it works a little bit like Spotify Wrapped, rounding up your activity in the app over the past month, three months, six months, or year. And it turns out that its summaries are mostly fair and insightful, based on my testing.
Reflect is available now inside the Claude apps for Free, Pro, and Max users, though Anthropic is labeling it as a beta feature for now.
Getting started with Claude Reflect
Before you can use Claude Reflect you need to have the Memory feature enabled: On the web, click your profile icon (bottom left), then Settings > Capabilities. In the mobile apps, tap the menu button (top left), then your profile icon, then Capabilities. If the Search and reference chats toggle switch is enabled, you're good to go.
When you're ready to do some reflecting, you need to be in the web or desktop apps. Click your profile icon (bottom left), then Settings, then Reflect. A summary will be automatically generated for your last month of usage, by default, but you can change the time period via the drop-down menu in the top right corner (click the circular refresh arrow icon to load a new summary).
I must admit up front that I tend to regularly delete my Claude chats as I go—I don't like leaving a long digital trail behind me for a variety of privacy and security reasons—so I can't tell you what it's like to have Claude dig through months and months of conversations. It may be that digging into bigger datasets get you different results.
What I can tell you is that Claude's review of my last month of AI use was still interesting: It correctly identified that I spend a lot of time working on little web apps and widgets that can help with my work, and of course experimenting with Claude's capabilities (also for work purposes, most of the time).
However, it does ignore your most recent conversations, depending on the calendar: If you're half way through July, your "past month" summary will only run up to the end of June, so only whole months are counted in the summary. It also seems to ignore shorter, one-off chats in favor of bigger trends.
Digging deeper into the data
Your reflection starts with a summary and a chart showing daily activity, but if you scroll down there's more to explore. Claude will break down your time into categories that you spend time on (like a New Tab extension for Chrome in my case), and assess your AI fluency skills too—basically, how good a prompter you are.
I scored highly in terms of "delegation" to the chatbot and "discernment" in terms of assessing the AI output, but apparently my "description" skills could use some work. Claude pointed out an occasion where I could've been more precise with a prompt, and gave me an example of how I could've refined it.
I'm not sure anything I read in Claude Reflect is going to change how I use the AI, but I think it's worth dipping into every now and again just as a self-audit. In the same way Spotify Wrapped can reveal that you're listening to far too much '90s pop, Reflect can help reassess your AI boundaries—particularly for heavy users.
I did find some of the summarizing to be a bit too polished: It's typical of AI bots like Claude to want to write a flowing narrative that can be tied off neatly, rather than a more awkward report that's more accurate, and this is evident again here. On the whole, though, it seemed to give me fair assessments.
If you do feel your Claude use has got out of hand, the Reflect summary includes a Set quiet hours and breaks link that leads to the Time and focus section of the app settings: From here you can set screen time limits for yourself with the AI, and have it remind you to take a break from prompting every so often.

LifeHacker

文章目录


    扫描二维码,在手机上阅读