我们用 AI 写邮件、改文案、做 PPT，却很少问自己：是不是从一开始，就把目标定得太小了？
这篇文章译自 Claude 开发者 Felix Rieseberg 的《How to be more ambitious: A tutorial for humans with models》。他分享的不是一套提示词技巧，而是三个值得尝试的方向：把目标提得更高，把问题挖得更深，让不同领域的方法碰撞出新的可能。如果你已经在用 AI，却总觉得它的价值不止于此，这篇文章或许能帮你打开思路。
作者：Felix Rieseberg 发表于：2026 年 10 月 13 日
我参与 Claude 的开发。这份工作的体验很有意思：每天，我都会听说有人借助 Claude，或由 Claude 自己，完成了某件了不起的事。有些事我从没想过，有些我以为不可能，还有些，只是我一直没有足够的好奇心和时间去尝试。就在这个月，它可能帮助人们发现了一颗新行星；我看到了第一支自己真心喜欢的 AI 音乐视频；我们还在讨论，数学是否正在进入一个新时代。
模型的各项能力并不是齐头并进的。正如陶哲轩所说，它们“在某些方面令人惊叹，在另一些方面却仍显不足”。有些人已经懂得驾驭那些出色的能力，仿佛真的飞了起来。看着人们用我参与打造的产品抵达惊人的高度，实在让人着迷。另一些人却还在摸索，怎样才能让这对新翅膀派上用场：除了问“法国的首都是哪里”，或者“能不能把这封邮件改得没那么冲”，它究竟还能做什么？
我和许多 AI 从业者经常说：使用 AI 时，我们应该更有雄心。这句话几乎已经成了口号。每当看到令人惊叹的成果，我们就转发，然后感叹一句：“我们的想象还是太保守了。”
但我也理解，真要照着这句话去做，并不容易。所以，我想写一份实践指南，聊聊怎样把目标定得更大胆。
在我看来，这种雄心可以体现在四个维度：速度、高度、深度和广度。
- 速度：用更少的时间，完成同样的工作。 这是 AI 创造经济价值的基本方式。比如，你需要写代码，我用 Claude 就能更快完成。这种用法已经很成熟，也不太需要你在使用方式上发挥创造力，所以这里就不展开了。
- 高度：提高思考的层次，关注最终成果，而不只是眼前的一项任务。
- 深度：像领域内的顶尖专家那样思考任务。 不停留在任务表面，而是再往下挖一层，深入底层的数据、工具和问题。
- 广度：借助多个领域的力量。 不局限于任务表面上所属的领域，而是用上你平时不会接触的知识、工具和方法。
我主要使用 Claude——这还用说吗——但这些建议并不只适用于某一个模型。
高度
我们常常随口把这叫作“让模型在更高的抽象层次上工作”。
当你给模型布置一项任务时，它执行的，其实是你目前认为最有可能实现目标的办法。想在高度上更进一步，就问问自己：接下来，我真正希望发生什么？眼前这个目标背后，还有什么更大的目标？
去年年初，我还在让 Claude 帮我修复一个个具体的崩溃问题。后来，我干脆让它负责让我们的应用保持稳定、不再崩溃。
来看一个日常的办公室任务——甚至可以说有点乏味。假设你在一家虚构的公司 Acme 工作，需要为季度业务复盘（QBR）制作一套幻灯片。给 Claude 一批数据，它就能很乐意地帮你做出来。也许你已经有了模板，它只需要往里面填内容。
这显然还谈不上发现新行星那样的壮举。那么，我们怎样一步步把目标放大？不妨从最基础的做法开始，逐渐往上走：
- 先弄清楚发生了什么。 大多数业务复盘想回答的，无非是：这个季度有什么变化，为什么会变，我们应该怎么办？别急着只更新模板里的数字。尽可能把你能拿到的数据交给模型，让它真正理解你的业务：“告诉我，这个季度发生了哪些变化，原因是什么，我们应该采取什么行动。”
- 进一步推动决策。 理想情况下，会议——至少是那些有价值的会议——结束时，应该能作出一些决定。你可以对 Claude 说：“根据你目前掌握的全部信息，找出最迫切需要作出的三个决策，开始梳理可能的行动方向，并分别列出支持和反对的理由。”
- 重新思考整个复盘机制。 季度业务复盘诞生于这样一个时代：要做好复盘，人们必须亲自仔细筛选数据，理解数据意味着什么，再判断业务是否需要调整。你可以对 Claude 说：“复盘的目的，是让管理层尽早发现问题，把资金和人力配置到合适的地方。请设计一种比季度汇报幻灯片更好的办法，用我们自己的历史数据证明，它本来可以更早发现问题，然后让这套办法与现有流程并行运行一个季度。”
当你真正追求的是某个最终结果时，提高目标层次最有用；对于创作类工作，它的适用性就没那么强。一个很好的判断信号是：如果你正在做的东西，之后还需要开会讨论，那么很可能值得把目标再往上提一层。
当然，目前仍然存在限制。到了 2026 年，单凭一句“Claude，去赚钱吧”，对大多数企业来说似乎还是行不通。不过，总体上我看到的是：模型能力越强，就越能在更高的抽象层次上开展工作。
人们也正是在这个维度上，找到能够长期运转、无需人工持续介入的工作流程。这样，你就不必不断给模型安排一个又一个新任务。
要做到这一点，往往需要改变自己的工作流程。比如，为了让 Claude 负责保障应用不崩溃，我得先确保它能方便地获取崩溃报告，知道每个应用版本对应哪一次 Git 提交，还要让它能够使用 Windows、macOS 和 Linux 电脑。只有这样，它才能真正独立复现问题，并验证修复是否有效。
深度
有时候，你只是想把一件事做到极致。
在深度上更有雄心，意味着推动模型采用这个领域最优秀的专业人士所使用的工具、技术、媒介和知识，而且继续往下钻研，直到连这些专家也可能从你的成果中学到新东西。
继续用 Acme 公司季度业务复盘的例子。假设这家公司正面临用户流失问题，也就是客户随着时间推移不断离开。我发现，结合使用下面这些方法很有帮助：
- 坚持采用业内顶尖人士使用的工具。 当你要求模型表现得像个专业人士时，它往往只是换了一套词汇，并没有真的换一种工作方法。要深入研究问题，我们就得使用顶尖数据科学家和统计学家采用的工具。当 Claude 准备拿 Markdown 表格来应付时，要坚持让它使用数据库和统计模型。你不一定非得知道专家用什么，Claude 通常知道。
- 定一个看起来不合理的目标，并告诉模型，这完全合理。 你原本只想把网站加载时间从 10 秒缩短到 5 秒？那就要求它做到 200 毫秒。如果模型觉得某件事不可能，就再次强调：这个问题完全可以解决，它只需要继续努力。
- 挑战你自以为知道的事情。 尤其是在商业领域，我们认定为真的许多判断，其实只是对现实的粗略近似。让模型接触原始信息和数据，要求它站在质疑者的立场检验假设。比如：“我们认为，用户流失是新用户引导环节出了问题。试着从原始数据中重新验证这个结论。”
- 发现新的知识。 这很容易接在检验既有认知之后：既然 Claude 已经深入数据，就让它继续寻找盲点。“根据我们的日志，构建一个客户行为模拟，重现过去一年的情况。然后找出我们尚未发现的、最早预示用户流失的信号，并用它没见过的数据证明，这个信号确实有预测能力。”
- 引入外部评判。 不要让模型自己给自己打分。我想，随着模型变得更聪明，这个问题最终会得到解决。但至少在 2026 年 10 月，我强烈建议你坚持使用外部验证。最好是某种模型能够自行运行、并立即获得反馈的工具。
下面这段简单的提示词，经常能帮到我：
我想把这项任务做得更深入，而不是把它的规模做得更大。在改写之前，请先回答四个问题：
- 这个领域的顶尖专业人士会做哪些事，而我目前没有提出这些要求？
- 他们会从哪些最原始的材料入手，而不是使用已经替我整理、加工过的信息？
- 他们具体会采用哪些方法和工具？
- 除了你我自己的判断，还有什么能够证明结果正确？
然后，请重写我的提示词，明确要求完成上述工作。
广度
简单来说，在广度上更有雄心，就是从不同领域的角度看待自己的工作，而且涉及的领域，可以比过去在成本上划算的范围更多。
这也不是什么独到的见解。哪部盗窃题材的电影，不是先花上半小时，召集一群各怀绝技的人？
随着 AI 产品让多智能体系统越来越容易运行——无论是通过多个会话、协调调度的机制，还是直接使用子智能体——我发现，反复要求 Claude 引入其他领域的方法，往往能带来很有意思的结果。
不过，把这件事落到实处并不容易，你得发挥创造力。直到看到同事 Alex Albert 用 Blender 渲染室内设计，我才想到，可以让 Claude 为我的个人主页制作三维场景。要知道，我以前甚至做过电子游戏！
但毫无疑问，这也是我最常看到惊喜的地方。
那么，具体该怎么做？我会采用下面这些方法：
- 试着把名词换成动词。 这是 Tom Kelley 和 Jonathan Littman 在《创新的艺术》中介绍过的老办法。季度复盘不是“一套幻灯片”，而是“在 30 分钟内，让一群忙碌的人相信一套关于因果关系的解释，并在不确定的情况下作出决策”。用动作来描述工作之后，我和 Claude 都更容易看清，哪些领域能够带来独特而有用的视角。
- 召集你的团队。 假设这套幻灯片必须精彩绝伦，否则外星人就要炸毁地球。人类会组建一支什么样的团队？为了保证成品质量，我会找一位分析师、一位记者和一位编辑。如果还要现场展示，可能还需要动态图形设计师和音频设计师。再找一支专门挑错、提出质疑的“红队”，也许同样有用，那就再请来一位做空投资者和一位法务会计师。
这里很容易一不小心就变成角色扮演：对话记录越来越奇怪，实际成果却没有变好。使用这个方法，是为了找到值得借鉴的领域和技术。别让模型模仿做空投资者的口吻写作，而要让它采用做空投资者研究投资标的时真正使用的方法。
我上面谈到的所有做法，都有一个前提：模型需要能够使用一台电脑，可以安装软件，也能够获取各种相关数据——无论是通过 MCP 连接器，还是其他途径。目前已经有不少工具能提供这些能力，我建议你选一个用起来。
类似的处境，我们以前也经历过
以上就是我关于如何把目标定得更大胆的具体建议。如果你愿意再读一会儿，我想解释一下：为什么我认为，我们首先需要认真对待这件事。
人类始终受到自己对“什么是可能的”的判断所限制。更麻烦的是，我们通常只在非常狭窄的专业范围内相信自己。
米开朗琪罗不愿意为西斯廷礼拜堂作画的故事，就很有名。毕竟，在他看来，自己是雕塑家，不是画家。
研究这个问题时，我发现心理学里有一个相关概念：自我效能感。简单来说，自我效能感强的人，更相信自己的能力，也因此确实能够完成更多事情。
有证据表明，我们可以通过一些方法提升自我效能感。Locke 和 Latham 梳理了长达 35 年的目标设定研究后，得出了一个明确结论：被赋予具体而有挑战性的目标的人，表现优于那些只是被要求“尽力而为”的人。
Stripe 创始人 Patrick Collison 一直维护着一份名为“Fast”的清单，记录人们在目标大胆、期限紧迫的情况下，完成过哪些了不起的事情。以下几种方法，都有充分的事例支持：
- 相信事情做得到，就可能完成难事。 1939 年，George Dantzig 在伯克利上统计课时迟到了。他抄下黑板上的两道题，以为是作业，解出来后交了上去。实际上，那是统计学中两道尚未解决的难题。
- 看到别人做到，自己也就更有可能做到。 几十年来，人们一直试图把一英里的跑步成绩突破到四分钟以内。Roger Bannister 在 1954 年 5 月做到了，但他作为唯一达成这一成绩的人，只保持了短短 46 天。如果你关注跑步，就会发现，全程马拉松突破两小时也正在上演类似的故事：今年 4 月，Sabastian Sawe 在伦敦跑出了首个获得正式认可的两小时以内的马拉松成绩。在同一场比赛中，亚军也突破了两小时。
- 允许自己失败，反而可能完成难事。 我上周才读到《网球的内在游戏》，结果发现，这似乎是人人都喜欢的一本商业书。它的核心观点是：如果你一门心思拼命想打好，你反而打不出自己最好的网球。我们这些工程师也许还记得，Linux 刚公布时，是这样介绍自己的：“只是个爱好，不会像 GNU 那样庞大而专业。”
自我效能感并不是新问题，但它如今的影响更大了。在一个我们的能力每隔几周就会跃升一次的世界里，我们需要一些方法，帮助自己去做那些从未认真考虑过、甚至不相信能够做到的事。
如果你发现了其他能让人更有雄心的方法，请告诉我！
本文配图为弗朗西斯科·戈雅的作品《飞行之道》（Modo de volar）。
原文链接：https://felixrieseberg.com/how-to-be-more-ambitious-a-tutorial-for-humans-with-models/
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
We write emails, rewrites, do PPTs, and rarely ask ourselves: Has the target been too small from the beginning?
The article is translated by Claude Developer Felix Rieseberg, How to be more ambitious: A tutuial for humans with mobiles. What he shared was not a set of prompt techniques, but three promising directions: raising the target higher, digging up the problem deeper and creating new possibilities for approaches in different fields. If you're already using AI, but you always think it's worth more than that, this article might help you open your mind.
Published by Felix Rieseberg: October 13, 2026
I'm involved in Claude's development. It's an interesting experience: every day, I hear people do something amazing through Claude, or by Claude himself. There are things I have never thought about, some I think it is impossible, and some I just haven't had enough curiosity and time to try. Just this month, it may have helped people discover a new planet; I saw the first AI music video that I really liked; and we're still discussing whether mathematics is entering a new era.
The capabilities of the model do not go hand in hand. They are “unsurprising in some respects, but still insufficient in others”, as Tao Chul-hyun said. Some already know how to handle those great abilities as if they really flew. It's fascinating to see people reaching amazing heights with products I'm involved in. Others are still trying to figure out how to make this work for the new wing: what else can it do, apart from asking “where is France's capital” or “can it change this mail so badly?”
I and many AI practitioners often say that when using AI, we should be more ambitious. This phrase has almost become a slogan. Every time we see an amazing result, we forward it, and we lament: “Our imagination is still too conservative.”
But I understand that it is not easy to follow that line. I would therefore like to write a guide to practice on how to set targets more ambitiously.
In my view, this ambition can be reflected in four dimensions: speed, height, depth and breadth.
- Speed: less time to do the same job. This is the basic way for AI to create economic value. For example, you need to write the code and I can do it faster with Claude. This use is mature and does not require your creativity in the way you use it, so it's not here.
- High: raising the level of reflection and focusing on the final outcome, not just the immediate task.
- Depth: Thinking about tasks like the best experts in the field. Instead of remaining on the surface of the mission, an additional layer of data, tools and problems go down to the bottom.
- Wideness: drawing on the power of multiple fields. It is not limited to what the mission appears to be, but rather uses knowledge, tools and methods that you do not normally have.
I'm mostly using Claude -- to be sure -- but these recommendations are not just for one model.
Height
We often call it "leting models work at a higher abstract level."
When you assign a task to the model, it carries out what you think is the most likely way to achieve it. If you want to go higher, ask yourself: what do I really want to do next? What is the bigger goal behind this?
Early last year, I was still getting Claude to help me fix a specific crash. Later, I simply made it the responsibility to keep our applications stable and not to collapse.
Looking at a day-to-day office assignment — even a bit boring, to say the least. Assuming you work for a fictional company, Acme, you need to produce a slide set for the QBR. Give Claude some data, and it'll help you make it. Maybe you've got a template, it just needs to fill in.
It is clear that nothing can be said about the magnitude of the discovery of new planets. So, how do we zoom in on the target? It would be useful to start with the most basic approach and move up gradually:
- Just figure out what happened. Most of the business rewinds want to answer the following question: What has changed this quarter, why has it changed, what should we do? Don't rush to just update the numbers in the template. Give the data you can get to the model as much as possible so that it truly understands your business: “Tell me what's changed this quarter, what's the reason, what we should do.”
- Further promotion of decision-making. Ideally, meetings — at least those of value — should end with decisions. You can say to Claude: “In the light of all the information that you have at hand, identify the three decisions that are most urgently needed, start combing the possible course of action and list the reasons for support and opposition, respectively.”
- Rethink the whole reset mechanism. The Quarterly Business Rover was born of an era in which people themselves had to carefully screen the data, understand what the data meant and determine whether the business needed to be adjusted. You can say to Claude: “The purpose of the CRT is to enable management to detect problems as early as possible and to place funds and manpower in the right place. Please devise a better way to report slides on a quarterly basis, proving with our own historical data that it could have detected problems earlier, and then allow the system to run a quarter in parallel with existing processes.”
When you really pursue a certain end result, it is most useful to raise the level of the goal; it is less relevant for creative work. A good judgement signal is that if you are doing something that needs to be discussed later, then it is likely that it will be worth taking the goal up.
There are, of course, still limitations. By 2026, a single phrase, “Claude, go make money”, still seemed unworkable for most businesses. In general, however, I see that the stronger the model, the better it is to work at a higher abstract level.
It is also within this dimension that people find workflow courses that can operate on a permanent basis without the need for sustained manual intervention. So you don't have to keep assigning the model a new task.
In order to do this, it is often necessary to change its own workflow programme. For example, in order for Claude to be responsible for ensuring that applications do not collapse, I have to make sure that it has easy access to crash reports, to know which Git submitted each application and to make it possible to use Windows, MacOS and Linux computers. Only in this way can it truly recover the problem of independence and verify the effectiveness of the repairs.
depth
Sometimes you just want to make one thing happen.
More ambitious in depth means that models are driven by the tools, technologies, media and knowledge used by the best professionals in the field, and continue to be developed downwards until even those specialists can learn from your achievements.
Continue to use the example of Acme ' s quarterly business round-ups. It is assumed that the company is facing a problem of loss of users, i.e., that clients are leaving over time. I find it helpful to combine the following methods:
- Maintaining the tools used by the top players in the industry. When you ask the model to act like a professional, it often simply changes a vocabulary and does not really change a working method. In-depth research requires the use of tools used by leading data scientists and statisticians. Claude insists on using databases and statistical models when it is prepared to use the Markdown table. You don't have to know what experts use, Claude usually knows.
- Set an unreasonable target and tell the model that it's perfectly reasonable. You were just trying to reduce the loading time from 10 seconds to 5 seconds? So it's required to do 200 milliseconds. If the model feels that something is not possible, it again stresses that the problem is perfectly achievable and that it needs only to continue.
- Challenge what you think you know. In the commercial sphere in particular, we believe that many judgements are true and that they are in fact mere approximations of reality. The model is exposed to raw information and data and is required to test assumptions in the opinion of the challenger. For example: “We think that the loss of users is a problem with the new user guide. Try to re-validate this conclusion from the original data.”
- Find new knowledge. It's easy to go after what's known: since Claude's in deep data, let him keep looking for blind spots. “By our log, a customer behaviour simulation is built to replicate the past year. Then we find the first signal that we have yet to discover and predict the loss of the user and prove with data it has never seen that the signal does have the ability to predict.”
- Introduction of external evaluation. Don't let the model score itself. I think that, as the model becomes smarter, it will eventually be resolved. But at least in October of 2026, I strongly recommend that you insist on external validation. It would be preferable for a model to operate on its own and obtain immediate feedback.
Here's a simple prompt that often helps me:
I want to do this task in greater depth, rather than on a larger scale. Before rewriting, please answer four questions:
- What would the best professionals in the field do, and I have not made these requests at this time?
- What are the most primitive materials they can start with, rather than using information that I have collated and processed?
- What methods and tools would they use?
- Is there anything other than your own judgment that can prove it right?
Then, please rewrite my prompt, with a clear request to complete this work.
Wideness
In short, there is greater ambition in terms of breadth, in terms of looking at one ' s work from a different perspective, and in terms of areas that can be more cost-effective than in the past.
It's not an opinion. Which one of the stolen films didn't take half an hour to bring together a bunch of people who could use their skills?
As the AI product makes the multi-AI agent system more and more operational — whether through multiple sessions, mechanisms to coordinate the movement, or directly using sub-AI agent — I find that repeated requests for Claude to introduce other fields often bring interesting results.
However, it is not easy to put this into practice, and you have to be creative. I didn't think until I saw my colleague Alex Albert replicating his interior design, that Claude could make a three-dimensional scene for my home page. You know, I used to play video games!
But there is no doubt that this is the place where I see surprises most often.
So, what exactly should we do? I'll do the following:
- Try changing the term to a verb. This is the old approach described by Tom Kelley and Jonathan Littleman in The Creative Arts. The quarterly flash drive is not a " set of slides ", but " in 30 minutes, to convince a busy group of people of a set of explanations for causality and to make decisions without certainty ". Now, with action to describe work, Claude and I can see more easily what fields can bring unique and useful perspectives.
- Call your team. Let's say this slide has to be brilliant, or the aliens will blow up the Earth. What kind of team would man form? To ensure the quality of the product, I will find an analyst, a journalist and an editor. There may also be a need for dynamic graphic designers and audio designers for live presentations. It might also be useful to find another “red team” to single out mistakes and challenge them, so that an empty investor and a legal accountant could be hired.
It is easy to become role-playing by accident: the record of dialogue is getting weirder and the actual results are not getting better. This approach is used to identify areas and technologies that deserve to be drawn upon. The model should not mimic the words of an empty investor, but rather use the methods that are actually used when an empty investor studies the investment mark.
All the approaches I've discussed have a premise: models need to be able to use a computer, install software, and access all the relevant data -- - Whether it's through the MCP connector, or otherwise. There are a number of tools available to provide these capabilities, and I suggest you choose one.
A similar situation. We've been through it before. Pass.
These are my specific recommendations on how to set targets more ambitiously. If you would like to read it again, I would like to explain why I think that we need to take it seriously first.
Humankind has always been limited by its own judgement of what is possible. Even more troublesome, we usually believe in ourselves only within very narrow professional boundaries.
The story of Michelangelo who refused to paint the Sistine Chapel is famous. After all, in his view he was a sculptor, not a painter.
When I looked at it, I found that there was a concept in psychology: a sense of self-effectiveness. Simply put, those who feel more self-efficient believe in their abilities and are thus truly able to do more.
There is evidence that we can improve our sense of self-effectiveness in a number of ways. Locke and Latham, after a 35-year target-setting study, came to the clear conclusion that those who were given specific and challenging goals performed better than those who were merely asked to “do their best”.
Stripe's founder Patrick Collison has been maintaining a list called "Fast" to document what great things people have done with ambitious goals and deadlines. The following approaches are well supported:
- If you can do something, you can do it. In 1939, George Dantzig was late for statistics class in Berkeley. He copied the two questions on the board, thought they were homework, and then handed them over. In fact, that is two of the outstanding problems in statistics.
- Seeing other people do it makes you more likely. For decades, people have tried to break a mile of running scores to within four minutes. Roger Bannister did it in May 1954, but as the only person who achieved it, he remained only 46 days short. If you're looking at running, you'll find that a similar story is going on for two hours: In April this year, Sabastian Sawe ran his first two-hour marathon in London. In the same game, the army also exceeded two hours.
- Allow yourself to fail, but it can be difficult. I read The Inline Game of Tennis just last week, and it turns out that it seems to be a business book that everyone likes. Its central point is that if you want to play well, you can't play your best tennis. And we engineers might remember that Linux, when he first published it, introduced himself like this: "It's just a hobby, not as big and professional as GNU."
Self-effectiveness is not a new issue, but its impact is greater today. In a world where our capacities will leapfrog every few weeks, we need ways to help ourselves do things that have never been seriously considered or even believed in being able to do.
If you find anything more ambitious, tell me!
This is illustrated by the work of Francisco Goya, Modo de volar.
Original link: https://felixrieseberg.com/how-to-be-more-ambitious-a-tutorian-for-humans-with-models/
Auto-fetched and translated by AI Atlas. All rights belong to the original author.