We Need to Be "More Ambitious" with AI

Most people using AI are limited by their own imagination, relying on it only to draft fluff emails or tweak formatting. What we really need is to be "more ambitious" with AI.

Felix Rieseberg, who contributes to Claude's development, frequently encounters uses of AI that he never anticipated. While he primarily uses Claude day-to-day, this advice is not tied to any specific model. Models demonstrate striking capabilities in certain directions while remaining distinctly lacking in others. These experiences convinced him that the demands we place on AI could stand to be far more ambitious.

He divides this ambition into four dimensions: Speed, Altitude, Depth, and Range.

What Can Change in the Same Task?

When preparing a quarterly business review, you can change your demands on AI in four directions:

1. Speed: Get the task at hand done faster.
   Tidying data, writing code, building slide decks—the objectives and steps remain essentially the same, but AI helps you finish sooner. Numbers you used to copy manually into a report template are now compiled and plugged in by the model, leaving you to verify the results.
   
2. Altitude: Look up to see what doing this is ultimately for.
   You have AI build a quarterly report to understand business issues and decide what to do next. Following that purpose, you can ask it to identify the three most urgent decisions, then design an early-warning process that catches operational issues sooner.
   
3. Depth: Dig into the problem like a top domain specialist.
   When analyzing why users churn, have it open the raw behavioral logs, use databases and statistical methods to test the hunch that "poor onboarding is to blame," and validate findings on unseen data. Trace down through data, tools, and key assumptions until the conclusions are grounded and hold up under scrutiny.
4. Range: Borrow signature methods from other fields.
   On that same quarterly report, you can verify figures using an analyst's methods, track down factual sources using a journalist's methods, and hunt for contradictory evidence using short-seller research techniques. Have these distinct professional disciplines tackle the problem to catch blind spots that a single perspective would miss.

These four directions can be applied individually or in combination. Tracing this quarterly business review shows how each type of demand can be taken further.

Speed: Letting AI Do the Same Work Faster

Work that previously required writing code, compiling spreadsheets, and crafting slide decks by hand can now be handed to AI to save time. This alone delivers real value, and it is the easiest entry point for most people.

Altitude: Handing Off a Larger Scope

When a manager asks for a quarterly review, you might instinctively tell AI to build a slide deck. Yet the ultimate purpose of this presentation is to uncover issues and make decisions.

So raise your demands: have it identify the quarter's most pressing issues, list the three most critical decisions, and evaluate the pros and cons of each direction. Take it a step further and have it design a routine workflow that catches anomalies on an ongoing basis—reducing the need to wait for the next quarterly review.

What you originally handed off was merely one step toward that objective.

Felix's own approach shifted from having Claude patch individual crashes to enlisting it to continually maintain application stability. To do that, he mapped out the relationships between crash reports, application versions, and code commits, then provided Windows, macOS, and Linux environments so the model could directly reproduce bugs and verify fixes.

Returning to the quarterly review: whether a new process actually works must also be verified. The original article recommends validating against historical records and running the new mechanism in parallel with the legacy process for a quarter; when reviewing history, use only information that was available at that point in time.

Elevating tasks in this manner suits work with clear, concrete objectives. If what matters to you is the personal craft of creation, there is no need to reflexively hand off the task.

Try It Out: Three Rounds of Prompts for a Quarterly Review (Chinese & English)

Here are the prompts adapted from the examples in the original piece. Supply relevant operational data first, then follow up round by round based on the analysis; choose either Chinese or English.

• Round 1 (Initial Request: Sent when preparing baseline review materials)
  Provide relevant operational data first to clarify what happened this quarter:

  Chinese comparison:
  “请结合我提供的完整经营数据，告诉我本季度业务发生了哪些核心变化、导致这些变化的原因是什么，以及我们接下来应该采取什么行动。”

  English version:
  "Based on the raw business data provided, tell me what key changes occurred this quarter, what caused them, and what concrete actions we should take next."

• Round 2 (Follow-up Request: Sent when debating decision directions)
  After reviewing the changes, ask which decisions demand immediate resolution:

  Chinese comparison:
  “根据你从数据中了解到的所有情况，梳理出摆在我们面前最紧迫的三项决策，并针对每项决策梳理出正反两方面的论证方案。”

  English version:
  "Based on everything you have learned from the data, identify the three most urgent decisions facing us, and outline potential directions with arguments for and against each."

• Round 3 (System Upgrade Follow-up: Sent when attempting to establish continuous monitoring)
  Previously, someone had to pore over data and conduct analyses to spot issues. Now you can have AI consider whether anomalies can be detected earlier to adjust staffing and capital in a timely manner:

  Chinese comparison:
  “业务复盘的存在，是为了让管理层及早察觉潜在隐患并将资金与人力调配到正确的方向。请基于我们公司的历史经营规律，设计一套比‘每季度做一次幻灯片’更及时有效的机制；请从历史数据中证明这套新机制能够更早拦截问题，并在下个季度与现有复盘流程并行运转验证。”

  English version:
  "Propose a process that helps leadership detect emerging problems and allocate resources sooner. Compare it with our quarterly presentation using historical records available at the time. Explain whether it would have warned us earlier, then plan a quarter-long trial alongside the existing process."

Depth: Digging into Problems Like a Top Domain Specialist

If a quarterly review reveals severe user churn, do not settle for generic platitudes like "optimize user onboarding" or "enhance customer care."

Direct it to inspect raw logs to see at which exact step users drop off, what actions they took beforehand, and how retained users differ. Query databases if queries are called for; run statistical analyses if statistical rigor is required. Once a pattern is identified, test it against unseen data to verify whether the finding holds up.

Writing "You are a senior expert" into a prompt is trivial. Having it actually work through a specialist's methodology is where genuine depth begins.

If the company has long assumed churn stems from flawed onboarding, instruct the model to verify this premise directly against the data, rather than hunting for evidence to confirm it. A signal that predicts churn can serve as an early warning; establishing it as the actual cause requires separate evidentiary backing.

Digging further into data can yield new insights: feed in a full year of raw logs, have the model construct a customer behavior simulation to uncover early churn indicators that the team had never recognized, and test their predictive validity on unseen data. The goal here is to unearth patterns that even seasoned practitioners might overlook.

For tasks like webpage acceleration, setting seemingly unreasonable stretch targets can likewise push the model to dig deeper. If a page currently takes 10 seconds and you plan to ask for 5 seconds, he suggests setting a target of 200 milliseconds to spur the model to isolate bottlenecks and test unconventional approaches. This serves as an exploratory target; whether it can actually be achieved ultimately hinges on real-world measurements.

Felix places special emphasis on independent validation outside the model's self-evaluations—ideally where the model can execute tools directly and receive immediate feedback. Run tests when writing code; evaluate performance on unseen data when making predictions; submit every result to verification.

Testing Established Assumptions, Then Finding New Signals

These two rounds of requests are adapted from the churn analysis example in the original article. Complete Round 1 before proceeding to Round 2; choose either Chinese or English.

Round 1: Testing the company's established assumption.

Chinese comparison:

公司一直认为，用户流失主要是新人引导造成的。
请从原始用户日志检查这个判断，列出支持它和与它冲突的证据。

English adaptation:

Our current explanation for churn is poor onboarding.
Check that explanation against the original user logs. Report evidence that supports it and evidence that conflicts with it.

Round 2: Looking for early indicators not previously noticed.

Chinese comparison:

请用过去一年的日志建立客户行为模拟。
先检查它能否复现已有记录中的变化。
再寻找我们此前没注意到、能更早预示流失的信号。
用未参与构建和选择信号的数据，检验预测是否仍然成立。

English adaptation:

Build a simulation of customer behavior using the last year of logs.
Check whether it reproduces the recorded patterns.
Look for early churn indicators we have not noticed before.
Evaluate their predictive value on data excluded from building the simulation and selecting the signals.

First, Reframe the Task at Hand with Four Questions

How to use: Tailor this to your own task. First provide your specific requirements and existing materials in the conversation, then send either of the following prompts (choose Chinese or English):

Chinese adaptation:

请帮我提高这项任务的专业深度，先不要扩大工作范围。
先回答以下四个问题：
1. 专业人员还会做哪些我没有提出的关键工作？
2. 他们需要哪些原始记录，而不只看现有摘要？
3. 具体应采用什么方法和工具，为什么适合这个问题？
4. 用什么独立证据或检验，可以判断结果是否可靠？
然后据此改写我的任务要求，写清执行步骤与验收方式。

English adaptation:

Help me improve the professional depth of this task without expanding its scope.
Before doing the work, answer four questions:
1. Which important steps would an experienced practitioner add?
2. What original records would they need beyond the summaries already supplied?
3. Which methods and tools fit this problem, and why?
4. What independent evidence or test would establish whether the result is reliable?
Use those answers to revise my task brief, including the work to perform and how to check it.

Range: Borrowing Expertise from Other Disciplines

Take that same quarterly report again. A financial analyst will cross-check the numbers, a journalist will track down sources, an editor will check for logical leaps in the argumentation, and a short-seller researcher will actively hunt for counter-evidence.

All of these methods can be put to work. Having AI verify the figures once, trace the sources once, and run a dedicated search for contradictory evidence is far more useful than slapping "expert" titles on several Agents and watching them nod in agreement.

One technique Felix uses is swapping nouns for verbs: instead of "make a slide deck," frame it as "help busy executives understand business shifts and make decisions in half an hour." Tracing these tangible actions makes it much easier to see which professional disciplines are missing. Verifying accounts might call for forensic accounting methods; delivering a live presentation might warrant motion graphics and audio design.

Cross-disciplinary execution is sometimes difficult simply because people never think to borrow other tools. Felix recounts having built games in the past, yet it was only after watching his colleague Alex Albert use Blender to render interior designs that he first thought to have Claude help construct a 3D environment for his personal homepage. The tools were right there all along; he simply had not conceived of using them that way.

Why Do We Often Hesitate to Demand More?

When confronted with new tools, people are often trapped within the boundaries of their existing habits. Felix draws on the psychological concept of "self-efficacy" to explain this mental barrier: people typically believe in their capabilities only within familiar domains, hesitating to set high goals in uncharted directions. He offers three suggestions: first, believe that the outcome is achievable; second, observe what others are already doing; third, give yourself permission to experiment and fail. As tools evolve, your judgment of "what I am capable of doing" can adapt alongside them. This addresses whether people are willing to try; whether the model's output is correct still requires verification.

To translate this analysis and verification into reality, the model must be able to operate computers, install software, and connect to relevant data via protocols like MCP, with runnable tools on hand to verify results immediately upon completion. Otherwise, being "more ambitious" easily degrades into nothing more than a grandiose prompt.