“Be more ambitious” is easy advice to give and difficult advice to use. Felix Rieseberg, who works on Claude, tries to make it concrete: stop treating better AI solely as a way to finish familiar tasks faster, and reconsider the outcomes, methods, and disciplines available to you.

The useful part is not permission to make a grander request. It is his account of what has to change around the model for that request to become meaningful. A model cannot keep an application reliable just because someone promotes it from bug fixer to reliability engineer. It needs evidence, tools, and a way to discover that its answer is wrong.

Four kinds of ambition

Rieseberg divides ambition into four dimensions:

  • Speed: accomplish the same work in less time. He treats this as the familiar case and largely sets it aside.
  • Altitude: move from the immediate assignment to the outcome it serves. Ask what should happen next, rather than accepting the current task as the right unit of work.
  • Depth: investigate with the raw material and methods a strong practitioner would use, rather than producing a more polished explanation at the surface.
  • Range: draw on additional disciplines when their techniques could improve the result.

These are different interventions. A longer report is not necessarily deeper. More agents are not necessarily more interdisciplinary. And a broader instruction is not necessarily a better delegation.

Altitude requires infrastructure

His clearest example is a shift from asking Claude to fix individual crashes to asking it to keep applications crash-free. That sounds like a prompting improvement until he describes the prerequisites: access to crash reports, a mapping between application versions and Git commits, and Windows, macOS, and Linux machines on which the agent can reproduce problems and verify fixes.

The larger outcome required a redesigned workflow. The agent needed to connect an observed failure to the relevant code and test whether its intervention worked. Without that loop, “keep the app crash-free” would be an aspiration disguised as an assignment.

His hypothetical quarterly business review follows the same progression. Filling a slide template is the low-altitude task. Explaining what changed and which decisions matter is a broader one. Designing an earlier-warning system, checking it against historical events, and running it alongside the existing review is broader still. That last step is a proposal, not a reported deployment, but its structure matters: a replacement process should demonstrate that it improves on the process it replaces.

Depth means changing the instruments

Rieseberg makes an unusually useful observation: models asked to act like professionals often change their vocabulary without changing their methods. A confident table about customer churn is not the same thing as an analysis of customer records.

His remedies are methodological:

  • Start from raw information instead of an already-digested account.
  • Use the tools the field actually employs, such as databases and statistical analysis, rather than settling for prose and tables.
  • Challenge the assumed explanation: if churn is blamed on onboarding, try to reproduce that claim from the records.
  • Look for previously unnoticed signals, then test them on data not used to discover them.
  • Use an outside judge—ideally a tool that supplies immediate feedback—rather than asking the model to certify its own result.

That last point is the load-bearing constraint. A model that proposes a theory and then declares it persuasive has not established that the theory predicts anything. A model that writes code and praises its implementation has not established that the program works.

His most reusable prompt begins:

“I want a deeper version of this task, not a bigger one.”

It then asks what a top professional would do, which raw material they would start from, which methods and tools they would use, and what—other than the model or user—would establish correctness. This is a better request than “think harder” because it asks for changes that can be inspected.

Range without the role-play

The interdisciplinary section also resists a familiar failure mode. Rieseberg proposes considering what an analyst, journalist, editor, forensic accountant, or adversarial researcher would contribute. But the point is not to make a simulated panel speak in different voices.

It is to borrow the work: checking records, reconstructing cause and effect, finding contradictory evidence, editing an argument, or testing a financial claim. Different personas using the same unsupported reasoning do not provide independent validation.

He suggests replacing nouns with verbs. A quarterly review is not simply “a deck”; it is an attempt to help people understand cause and effect and decide under uncertainty. Describing the activity makes it easier to identify which other disciplines have useful methods. He also describes discovering agent-assisted Blender work through a colleague—an example of expanding the available toolset, rather than merely expanding the prompt.

Where the advice needs a brake

Not every recommendation is equally sound. Rieseberg suggests choosing unreasonable targets and reassuring the model that they are achievable. Stretch goals can uncover approaches that a modest request would never explore. But pressure is not evidence. Insisting on an impossible result can also produce invented success, weakened acceptance criteria, or an expensive loop with no useful progress.

A safer formulation is: set a demanding target, require measurements, and allow the agent to report a real limit. Persistence should deepen the investigation, not suppress an honest failure report.

There is also an important distinction between the essay’s argument and its evidence. The opening capability anecdotes and closing stories about human self-efficacy do not establish the reliability of autonomous model workflows. His professional affiliation supplies practical experience and a commercial stake. This is a practitioner’s framework, not a controlled demonstration that broader delegation works across domains.

The useful ambition

The strongest reading of this essay is not “trust the model with more.” It is “build a better way to ask, investigate, and check.”

Raise the level of the outcome where that makes sense. Deepen the methods rather than inflating the deliverable. Bring in unfamiliar tools when they solve a real problem. But keep the result answerable to something outside the conversation: a reproduced failure, a working artifact, a prediction on unseen data, or an explicit human acceptance test.

Ambition earns its keep when it changes what can be demonstrated—not merely what the agent is willing to claim.