Anthropic Engineer's Framework for Ambitious AI Usage: Altitude, Depth, Range

Introduction: Beyond Speed — The Ambition Gap

Anthropic engineering lead Felix Rieseberg (formerly of Electron, Notion, Stripe, Slack, Microsoft) observes that while LLMs like Claude exhibit astonishing capabilities in some directions, they remain clumsy in others. Most users treat them as faster typewriters — asking for email polish or capital lookup — equivalent to "flying a J-20 fighter to deliver takeout." Rieseberg urges users to "Be More Ambitious" across three dimensions: Altitude, Depth, and Range.

Altitude: Raise the Abstraction Level

Instead of micromanaging steps, define the ultimate outcome. Rieseberg illustrates with his own shift from asking Claude to fix individual app crashes to commanding: "Keep our client application at zero crashes continuously." This triggered a fully automated pipeline: Claude pulls crash reports, maps versions to Git commits, spins up Windows/macOS/Linux test environments, reproduces bugs, verifies fixes, submits patches, and runs regression tests autonomously.

A QBR (Quarterly Business Review) ladder demonstrates the progression:

Step 1 (Basic): Manually fill a PPT template; maybe ask AI to polish wording.

Step 2 (Ambitious): Feed all raw business data to the model; ask: "Tell me what changed this quarter, why, and what actions we should take. Surface the top 3 urgent decisions with pro/con arguments."

Step 3 (Ultimate): Re-examine the purpose of QBRs — early problem detection and resource allocation. Prompt: "Design a new mechanism that beats quarterly PPTs at early detection. Prove it with historical data and run it in parallel next quarter."

Technical writer Thorsten Ball publicly praised this case study. Rieseberg offers a heuristic: if your work still requires frequent meetings to discuss and decide, it's a candidate for Altitude reframing.

Depth: Force Expert-Grade Execution

Merely prompting "You are a senior architect" yields only jargon, not expert methods. Rieseberg prescribes five tactics:

Demand real expert tools: Reject Markdown tables; require industrial databases, rigorous statistical regression, causal inference. If unsure, ask the model what experts use.

Set seemingly impossible targets: Don't ask to cut load time from 10s to 5s; demand 200ms. When the model balks, insist: "This is engineering-solvable; keep digging."

Use raw data to challenge assumptions: Feed original logs: "Management thinks churn is due to poor onboarding. Prove or disprove using only raw logs."

Create novel knowledge: Build micro-simulators from logs to predict unseen data, uncovering early churn signals veterans missed.

Introduce objective external judges: Never let the model self-evaluate. Require runnable code, unit tests, benchmarks — objective feedback loops for rapid iteration.

Rieseberg provides a reusable "Golden Depth Prompt" (translated below) that forces the model to articulate expert methods, raw data sources, professional tools, and external validation criteria before rewriting the user's prompt.

I want to make this task deeper, not just larger.

Before rewriting my prompt, do four things:
1. Identify what top experts in this field would do that I haven't considered.
2. Specify the rawest source materials and data they would start from, discarding second-hand conclusions.
3. List the concrete professional tools, statistical models, and working methods they would employ.
4. Define what objective external standard or tool — beyond you and me — can prove the result is completely correct.

After clarifying the above, rewrite my prompt to match that top-expert standard.

Range: Assemble Cross-Disciplinary Agent Teams

Modern multi-agent systems make cross-domain expertise cheap. Rieseberg borrows two tricks from IDEO's "Art of Innovation":

Swap nouns for verbs: Instead of "make a QBR deck" (noun), decompose into verbs: "In 30 minutes, make a group of busy, distracted people grasp causal relationships and dare to make major decisions under uncertainty." This verb framing instantly reveals needed perspectives.

Recruit a "heist movie" specialist crew: Data Analyst (cleaning, correlation), Investigative Journalist (human stories behind numbers), Editor-in-Chief (narrative arc, cut fluff), Motion/Audio Designer (pacing, visual hooks), Short-Seller Analyst (adversarial audit for flaws), Forensic Accountant (penetrate financial veneer).

Crucially, don't just role-play ("you are a short-seller") — that yields slick talk. Instead, require the model to invoke the actual research methods and audit steps those professionals use.

Breaking Through: Self-Efficacy and Historical Proof

Rieseberg explores why we hesitate despite powerful tools. Psychology calls it "self-efficacy" — belief in one's capacity to execute. Three historical cases show how shattered mental limits unlock breakthroughs:

George Dantzig (1939): Mistook two famous unsolved statistics problems for homework; solved both because he didn't know they were "impossible."

Roger Bannister (1954): Broke the 4-minute mile, deemed physiologically impossible. Within 46 days, a second runner also broke it — the barrier was psychological.

Linus Torvalds: Launched Linux as a "hobby, won't be big and professional like GNU." Shedding perfectionism enabled the greatest OS project.

Independent developer @levelsio noted Felix often voices high-level instructions to Claude while cycling, producing stunning 3D personal sites — demonstrating low-friction, high-altitude delegation.

Real-World Pushback: The Costs of Ambition

The article sparked vigorous debate among practitioners:

Creative serendipity risk: Screenwriter Alex Dobrenko warned that artistic breakthroughs often come from aimless wandering; excessive Altitude may kill spontaneity. Felix agreed but noted Altitude still frees creators from formatting drudgery to focus on story core.

Review cost explosion: Engineer Rajesh observed that high-abstraction tasks cause exponential growth in human verification effort. If humans cannot understand or validate the complex system, ambition becomes catastrophe.

Error snowballing: Developer JohnKocdr cautioned that tiny early errors in long-horizon tasks amplify catastrophically over dozens of steps.

These critiques reinforce a core principle: Ambition ≠ abdication. High-altitude tasks demand rigorous automated testing, version control, and objective external verification sandboxes.

Actionable Takeaways

Escape the "speed trap"; operate at Altitude by questioning the ultimate outcome.

Reject shallow role-play; enforce Depth with expert tools, raw data, and external validators.

Transcend single-discipline thinking; apply Range by verbifying nouns and assembling cross-functional agent teams.

Equip models with automated, verifiable execution environments; stay vigilant about review costs and error accumulation.

Pick your most painful complex task today, apply the Golden Depth Prompt, and witness the step-change in output quality.