You've probably caught yourself doing this: you open Claude, ask it to polish an email paragraph, fix a typo, or sketch out a quick helper function, and walk away satisfied. Stop and think about it for a second, though—you're sitting in front of one of the most capable computing systems ever built, and using it like a fancy spellchecker.
Anthropic lead engineer Felix Rieseberg—who oversees Claude Desktop and Claude Cowork—published an essay titled "How to be more ambitious: A tutorial for humans with models." In it, he dissects a psychological bottleneck across the industry that he calls Model Overhang.
It's the massive gap between what frontier models (Claude 3.5 Sonnet, Opus 5.5) can actually pull off today and the tiny, safe chores people timidly hand them. Let's look at how Rieseberg suggests closing that gap.
Four Dimensions of Ambition
Rieseberg lays out a simple mental model for scaling how you work with AI. Most users get stuck on rung one without realizing the other three even exist:
- Speed. The baseline tier: "do what I already know how to do, just in three seconds instead of half an hour." Boilerplate scripts, article summaries—handy, sure, but it doesn't fundamentally change your output.
- Altitude. Moving from micro-instructions to end-to-end goals. Instead of "fix this NullPointer bug," you hand over the whole arc: "analyze the last 24 hours of production logs, isolate the memory leak, write a standalone regression test, and open a minimal Pull Request."
- Depth. Giving the model the authority and context of a deep specialist. Hand it raw database dumps, performance profilers, and protocol specs so it can run a multi-step audit that would take a human analyst weeks.
- Range. Pulling in non-obvious methods from completely different fields—like borrowing population-dynamics math from biology to tune cache eviction in a high-load database.
The Iron Rule: Never Let a Model Grade Its Own Work
The most common rookie mistake in prompting is asking the same chat session to review its own output: "Now check the code you just wrote for bugs."
Rieseberg is blunt about this: it triggers severe confirmation bias. The model will almost always rationalize its own hallucinations and insist the code is spotless.
Never let a model judge its own work inside the same session. Verification must be strictly external and deterministic: run it through a compiler, execute automated tests in the terminal, or pass the artifact to a completely isolated reviewer in a fresh context.
Environment Is Everything
Getting real leverage from Claude isn't about hunting for "magic words" in your prompt—it's about wiring up a capable tool environment. A model hits its stride when it has hands: access to the filesystem, local test runners, a terminal, and external APIs via the Model Context Protocol (MCP).
Earlier, we covered collaborative work on Claude Artifacts and ready-to-use skills and repos for Claude. Rieseberg's guide drives home the core lesson: stop treating the AI like a chat buddy, and start treating it like an autonomous engineer backed by real tools and unforgiving tests.

