Every day, an engineer at Anthropic watches people do extraordinary things with Claude. This month alone, someone may have used it to help spot a new planet. He saw the first AI-generated music video he genuinely loved. Terence Tao — arguably the greatest mathematician alive — is openly discussing whether mathematics has entered a new era because of AI.

And then he watches the rest of us ask: "What's the capital of France?" and "Can you make this email sound nicer?"

Same model. Wildly different ambition. So he did what engineers do: he wrote a tutorial.

His core argument is simple — you should be more ambitious with AI. But he knows slogans don't change behavior, so he breaks ambition into four levels. The first is speed: doing the same work, faster. Everyone's already there. It's the other three where most people are leaving enormous value on the table.

Here's the full playbook.

Level 1: Raise the Altitude

When you give AI a task, you're usually asking it to execute your best guess at the right solution. The trick is to climb one level higher and ask for the outcome instead.

The engineer's own story is the perfect example. Early last year, he was using Claude to fix specific crashes in his code — one bug at a time. Then he changed the instruction entirely: "Make sure our app doesn't crash."

One is a bug fix. The other is system reliability. The moment the goal moved up a level, what the AI could do for him expanded dramatically.

The quarterly report example

Imagine you need to build your company's quarterly review deck:

That fourth level is the point: the quarterly report only exists because humans were too slow to analyze data continuously. Once AI makes continuous analysis cheap, the artifact itself becomes the bottleneck.

A quick test he offers: if you'll need a meeting to discuss the output anyway, you probably aimed too low.

The catch: higher-altitude work requires giving AI the infrastructure to match. To have Claude keep an app crash-free, he first gave it access to crash reports, a way to map versions to code, and a machine where it could reproduce bugs and verify its own fixes. An AI asked to be responsible for an outcome needs the tools to actually see and touch that outcome.

Level 2: Go Deeper

Sometimes you don't need a bigger goal — you need the same goal done to world-class standard. That's when you ask AI to work like the top expert in the field would, using the tools, methods, and raw materials that expert would actually use.

But there's a trap: tell AI to "act like a data scientist" and it often just adopts fancier vocabulary while doing the same shallow work. Ask it to analyze churn and it'll hand you a table — while a real data scientist pulls a database, builds a statistical model, and stress-tests assumptions. You have to specify the actual expert workflow, not the costume.

His techniques for going deeper:

• Set an absurd target. Website loads in 10 seconds? Tell the AI you want 200 milliseconds. Unreasonable goals force it past conventional thinking into territory your normal prompts never reach.
• Attack your own sacred assumptions. Company lore says churn is an onboarding problem? Hand the AI the raw data and ask it to check whether that belief actually survives contact with reality.
• Hunt blind spots. Have it build a simulation model of user behavior from historical data and surface the earliest possible churn signals — things you'd never think to look for.
• Never let AI grade its own homework. Self-evaluation by current models is unreliable. Validate results with external tools and methods.

The prompt he actually uses

"Before you rewrite my task: tell me how the world's top experts in this field would approach it — what raw materials they'd start from, what tools and methods they'd use, and what could judge the result besides you. Then rewrite my task to those standards."

That's it. Paste it before your next complex request and watch the quality jump.

Level 3: Build a Cross-Functional Team

The final level: look at your problem through the eyes of multiple disciplines at once. Historically this was impossibly expensive — you can't hire a panel of experts to review every quarterly deck. AI makes it nearly free.

His trick for assembling the team: convert the task from a noun into a verb.

"Quarterly report" is a noun — a thing with a template. Now describe it as an action: "Convince a room of busy people of a causal story in 30 minutes so they can make a decision under uncertainty." Suddenly you can see which fields to borrow from: an analyst for the data, a journalist for the narrative, an editor for quality control. Presenting on stage? Add a motion designer and a sound designer. Want it bulletproof? Add a short-seller-style researcher and a legal auditor whose entire job is to find the holes.

The pitfall here is roleplay theater. If the "short-seller" just writes paragraphs in a skeptical tone, you've gained nothing. The instruction isn't "write like a short-seller" — it's "do the research work a short-seller would do": dig for counter-evidence, model the downside, check the numbers that would embarrass us.

All three levels share one requirement: the AI needs a machine with real software installed, and access to your actual data. Ambition without infrastructure is just a fancier chat.

The Part Nobody Expected: A Lesson in Psychology

The essay ends somewhere surprising — not with prompting tricks, but with self-efficacy: the psychological finding that the more you believe you can do something, the more you actually can. And self-efficacy, it turns out, can be hacked.

Consider the evidence he lines up:

• 1939: George Dantzig arrived late to class, copied two problems off the board assuming they were homework, and solved them. They were famous unsolved problems in statistics. He solved them because nobody told him they were impossible.
• 1954: Roger Bannister broke the four-minute mile — a barrier deemed physically impossible for decades. Forty-six days later, someone else did it too. Once it was thinkable, it was repeatable.
• April 26, 2026: Sebastian Sawe ran the London Marathon in 1:59:30 — the first officially recognized sub-two-hour marathon in history — and Yomif Kejelcha crossed the line 11 seconds behind him, also under two hours, in the same race. Both said they could go faster.
• The Inner Game of Tennis: the harder you try to hit the ball well, the worse you play.
• 1991: Linus Torvalds announced Linux as "just a hobby, won't be big and professional like GNU." It now runs most of the world's servers.
• Michelangelo didn't want to paint the Sistine Chapel — he considered himself a sculptor, and painting "wasn't his job." The boundary was in his head, not his hands.

The engineer's closing point lands quietly but hard: AI capabilities improve every few weeks; our imagination about what we can do with them updates far more slowly. When Terence Tao — a Fields Medalist — says AI is changing how mathematics itself is practiced, as an "intellectual sparring partner" for exploring ideas before rigor begins, the barrier clearly isn't the tool anymore. It's the mental model of the person holding it.

The Takeaway

The gap between "fix this bug" and "make our app unbreakable" isn't a bigger model or a more expensive plan. It's a one-line change in how you frame the request — plus the willingness to give the AI real tools, real data, and an audacious target.

The engineers discovering planets and reinventing math aren't using a different Claude. They're asking a different question.

What's the highest-altitude version of the task in front of you right now? Drop it in the comments — the best answers usually sound impossible until someone does them.

Frequently Asked Questions (FAQ)

Who wrote the original essay this post is based on?

An engineer at Anthropic who works on Claude full-time, writing about how he uses the tool he builds. The essay circulated widely in both English and Chinese tech communities in late 2026.

What does "raising the altitude" mean in AI prompting?

Replacing a task-level request ("fix this crash") with an outcome-level request ("keep our app from crashing"), which fundamentally expands what the AI will attempt. It requires giving the AI access to the reports, data, and environments needed to actually own the outcome.

Why shouldn't I let AI evaluate its own work?

Current models are unreliable self-evaluators — they tend to grade their own outputs favorably. Always validate important results with external tools, data, or independent checks.

What's the noun-to-verb trick for AI teams?

Describe your task as an action rather than an artifact: not "a PPT" but "convince busy executives of a causal story in 30 minutes." Verbs reveal which expert disciplines (analyst, journalist, editor, auditor) can contribute — then have each one do the actual work, not roleplay the persona.

Is this relevant if I don't code?

Yes — the quarterly-report and presentation examples are pure office work. The framework (altitude, depth, cross-functional perspectives) applies to marketing, analysis, writing, operations, and research equally.