写给用模型的人类：怎样使用AI更有野心

To Humans Using Models: How to Use AI More Ambitiously

According to Felix Rieseberg, the engineering lead at Anthropic, as AI gets smarter, this usage pattern can no longer keep up with the models. When you tell the model to “act like an expert,” it usually just swaps in more professional terminology, but the output remains the same.

He has overseen the engineering of Claude.ai, Cowork, and the Claude Code desktop app, and almost every day he sees someone using Claude to do something he never imagined. Yet many people still use it to ask “What is the capital of France?” or to tone down an email “so it’s not so aggressive.”

So he wrote a tutorial on his personal blog titled “How to be more ambitious: A tutorial for humans with models,” which means “A tutorial for humans using models: How to be more ambitious.”

Article URL:

https://felixrieseberg.com/how-to-be-more-ambitious-a-tutorial-for-humans-with-models/

“Ambition” is a buzzword in the tech‑innovation circle. When AI practitioners see someone achieve something amazing with a model, they share it with a caption: “We’re not ambitious enough yet.” Felix says this too, but he knows it’s easy to say and hard to actually do.

He breaks “ambition” into four dimensions:Speed, Height, Depth, Breadth.

Speed means doing the same work faster, which is the most mature use of AI, and he skips it. The remaining three are what he really wants to discuss.

◈1. Height: From completing a task to clarifying the real goal

The first direction he mentions is “height.” In simple terms, when assigning a task to AI, ask one level higher: why am I doing this?

A question Felix raises in the article nicely explains his point:

“What do I really want to happen next? Is there a deeper goal behind the current objective?”

To help readers understand better, Felix also gives an example:

Suppose you are an employee, and your boss asks you to create a quarterly business review PPT. How would you use AI to accomplish it?

According to our usual habit of using AI, we might hand the data over to Claude, have it organize it, and then automatically generate a nice PPT.

Reaching this step already saves a lot of time.

But the author believes this is still not enough.

The purpose isn’t just to create a PPT; the real goal is to let the company know what happened this quarter, why it happened, and what to do next.

Therefore, instead of having AI merely update the numbers in your PPT, let it analyze the company’s operational data, identify issues, and even suggest the next decisions.

Taking it a step further, if AI can continuously analyze business data and spot problems in real time, does the company really need to wait until the end of each quarter to hold a separate review meeting?

So Felix suggests letting AI think about whether, besides doing a PPT report each quarter, there’s a way to spot business issues earlier?

For example, we could have AI analyze the company’s historical data to see if any overlooked early‑warning signals can be found. If those signals prove useful, we could try a new method to monitor the next quarter’s performance and compare it with the traditional quarterly report to see which approach catches problems more promptly.

I find this reminder extremely practical.

Often, when we give AI a task, we’re just transplanting our old work habits onto it unchanged.

What used to require us to gather data ourselves now we ask AI to gather; what used to require us to build spreadsheets now we ask AI to build them. Even if the original workflow was inefficient, we still make AI repeat that same process.

Of course, this isn’t to say that every future task must be redesigned by AI first.

It’s just that next time we face a complex, repetitive task, we might ask an extra question: besides helping me finish it, could AI offer a better approach?

◈II. Depth: From Playing the Expert to Using Professional Methods

The next step is “depth,” which means getting AI to go beyond the surface of the problem and use the tools and methods of professionals to examine something more thoroughly.

These days many people like to add to their prompts: “You are an industry expert with 20 years of experience.”

But the question is, does doing this really make AI more professional?

Felix believes that when you ask a model to “work like a professional,” it usually just changes the wording, uses more jargon, but doesn’t necessarily change the way it works.

For example, a company recently lost a lot of users. You ask the AI, how should this problem be solved?

It might tell you that you need to improve user experience, optimize product features, and increase user satisfaction.

These statements are certainly correct, but think about it: almost any company facing user churn could receive similar advice. The real reasons why users leave and which problems should be prioritized remain unanswered.

Felix suggests, instead of having AI play the role of a professional data analyst, let it use the tools that real data analysts actually use.

Let’s take this fictional company as an example and assume that internally it has always believed that user churn is due to inadequate onboarding for new users, causing many registrants to not know how to use the product.

According to past practice, employees might tell the AI this judgment and then ask it to suggest some optimizations, such as simplifying the registration process, adding a beginner tutorial, and so on.

But the author suggests doing the opposite: give the raw data to the AI and let it try to reproduce this conclusion to see whether it’s actually correct.

For example, you could let the AI access the company’s existing user behavior data, including when users register, which features they use, how often they log in, and when they eventually stop using the product.

Next, ask it to use databases and statistical analysis tools; if the data shows that a large number of users indeed leave within the first few days after registration, then new‑user onboarding may be a direction worth further investigation.

But if churn mainly occurs after a few months of use, the issue may not lie in the registration process, and further examination of product features, usage frequency, pricing, or other factors is needed.

In this way, AI no longer just follows our judgment to provide solutions, but helps us verify whether the presumed problem actually exists.

Felix also shared a set of prompts he frequently uses, which can be copied and used directly:

Compared to simply saying “you are an industry expert,” this is obviously much more specific.

◈3. Breadth: From a Single Perspective to Cross‑Disciplinary Solutions

The third type is called breadth: looking at the same task from more professional fields.

Felix mentioned that in the past, completing a complex job often required gathering a group of people with different backgrounds. Because their professional backgrounds differ, their perspectives on problems and solutions can be completely different.

For example, a company discovers a serious recent user churn and wants to understand the cause. If only a data analyst works on it, they might produce a detailed data report, but it may not explain why users are leaving. If only a journalist interviews users, the limited sample might not reveal how many people are affected by a particular issue.

In the past, getting them to analyze together was too costly, but now with AI we can leverage it to combine knowledge and methods from different fields.

The question is, how do we determine which experts we need and how to get them to help?

At this point, we can first use a data analyst’s approach, letting AI segment data by registration time, user type, and other dimensions, analyze retention rates of different groups, and identify when users tend to churn.

But those numbers may not tell us why users leave.

So we can bring in the methods of investigative journalists. Have AI design interview outlines based on existing data, compile customer service complaints, compare feedback from different users, and help us uncover the reasons behind the data. For example, is it because the product is hard to use, the price is too high, or users have found a better alternative?

If you find that users frequently abandon a particular step, you can also use the product designer’s approach to map out the steps required to complete the action, identify confusing interfaces and flows, and propose improvement plans that need further testing.

Finally, you can borrow the method of short‑selling analysts to conduct a stricter review of the conclusions drawn earlier: Are we only emphasizing active user growth while overlooking the churn of high‑value users? Some metrics look good—could that be just because the measurement criteria have changed? Is there any evidence that could overturn our current explanation for user churn?

Here Felix specifically reminds us: don’t let AI play role‑playing. It’s not about having several AIs act as different experts and pretend to hold a meeting, but about having AI actually use the work methods of those professional fields.

Here Felix specifically reminds us: Don’t let AI role‑play. Don’t have it “write like a short‑seller analyst”; instead have it conduct the actual investigations that a short‑seller analyst would do. The former only makes the conversation odd, while the latter can change the outcome.

This is actually appealing to ordinary people, because most of the time we lack perspective rather than effort.

In the past, when encountering these problems, you either had to learn slowly on your own or hire a professional, both of which were costly.

Of course, this does not mean AI can truly replace professionals in various fields.Its more realistic role is to help us identify our blind spots earlier, preventing us from framing the problem too narrowly from the start.

◈IV. Finally: From Trusting Answers to Verifying Results

In fact, this article says a lot,but the core idea is simple: encourage everyone to make bolder demands of AI.

He gives an extreme example: a web page now takes 10 seconds to load, you want it down to 5 seconds? Then just ask for 0.2 seconds. If the model says it can’t, tell it it can be solved completely and keep pushing.

This line of thinking is interesting, but it doesn’t mean AI can do everything.

Felix himself admits that by 2026, simply telling AI to “go make money” won’t get most companies moving. Moreover, all these approaches assume a premise: the AI must have a computer that can install software and access real data.

Even more importantly, AI sometimes solemnly tells us that it has completed a task, when in reality it may have only generated an answer that looks good.

So the author also emphasizes:“Don’t let the model judge its own work.”

For example, the same company dealing with user churn mentioned earlier.

If AI, by analyzing historical data, finds that certain user behaviors are linked to churn and even builds a model to predict churn, we can certainly ask it to refine the analysis, but we shouldn’t assume it’s truly effective just because the AI thinks the model performs well.

A more reliable approach is to take a portion of data that was not involved in the original analysis and test whether it can accurately predict those users’ behavior.

Other tasks are the same. If you ask AI to write code, you must actually run tests; if you ask AI to do research, you need to go back to the original sources to verify the conclusions; if you ask AI to produce a business report, you also need to check the data sources and calculation process, not just let it judge whether the report is professional.

In other words, we not only need to tell AI what task to accomplish, but also think ahead about what the standards for completing the task are.

This is actually a capability required for using AI more boldly.

After all, when we only ask AI to polish a piece of text, we can usually judge the result ourselves. But if we ask it to analyze thousands of data points or to complete a specialized task we are not familiar with, we cannot rely solely on our intuition, nor can we fully trust AI’s self‑assessment.

Therefore, daring to set higher demands on AI and finding reliable ways to verify the results are what enable it to truly take on more complex tasks.

In the end, the core of this guide is not to teach us how to write longer, more complex prompts, but to provide a more concrete set of methods for using AI.

From speeding up work to rethinking goals, delving into problems, and borrowing methods from other fields, what Felix wants to encourage is that we stop treating AI merely as a tool for efficiency and instead try to use it to tackle complex tasks we previously didn’t know how to approach.

Next time you use AI, perhaps start with a simple question and try his suggestion: besides doing it the way I originally intended, is there a better way?

References

[1] Felix Rieseberg: https://felixrieseberg.com/how-to-be-more-ambitious-a-tutorial-for-humans-with-models/