Prompt Engineering Isn’t Dead — You’re Just Focusing on the Wrong Half

Andrej Karpathy, one of the most respected voices in AI, recently made a statement that sent ripples through the community: investing heavily in prompt engineering isn’t worth it. Prompts, he suggested, will become devalued as models improve.

At first glance, this feels almost ironic. Karpathy’s own tweets and explanations are routinely turned into ready-made prompt templates and “lifehacks” shared across the internet. Yet his logic is sound once you separate the two very different activities that hide under the single umbrella term “prompt engineering.”

The Two Halves of Prompt Work

The first half is what most people mean when they talk about prompt engineering today: clever phrasing tricks, role-playing instructions (“You are an expert in X”), chain-of-thought prompts (“think step by step”), and other incantations designed to squeeze better performance out of a model. This layer is largely a consumable. As large language models get smarter and more robust, they become increasingly forgiving of messy or suboptimal inputs. Polishing these “magic phrases” is like optimizing for a moving target that is rapidly getting easier to hit. The ROI on mastering every new trick declines quickly.

Prompt Engineering Isn’t Dead — You’re Just Focusing on the Wrong Half The second half is far more valuable and durable. It involves breaking down complex, real-world problems into manageable pieces, deciding what to hand off to AI versus what to handle yourself, knowing what to verify, how to iterate productively, and how to synthesize the outputs into a coherent result.

This is less about wording and more about orchestration and judgment — what some are now calling “context engineering” or, more broadly, “AI piloting.”

Karpathy captures the distinction elegantly: You can outsource thinking to AI, but you cannot outsource understanding.

AI systems are already excellent (and getting better) at generating drafts, exploring variations, writing code, and iterating quickly. What they still lack — and what humans uniquely provide — is the ability to maintain the big picture, recognize what “good” actually looks like in a specific domain, steer toward meaningful goals, and apply hard-won expertise to evaluate outputs critically.

Why This Matters Now

Prompt Engineering Isn’t Dead — You’re Just Focusing on the Wrong Half There is no magic AI pill that will let anyone achieve expert-level results with zero domain knowledge. Much of what gets marketed as “just prompt better” is commoditized and will only become more so. The lasting advantage comes from genuine understanding — the kind built through real expertise, experience, and deliberate exposure to a field.

This perspective feels especially grounded coming from Karpathy. He was a founding member of OpenAI, served as Director of AI at Tesla (leading the computer vision efforts for Autopilot), returned to OpenAI for a period, and in May 2026 joined Anthropic to work on its pre-training team — specifically helping use Claude to accelerate pre-training research itself.

His background gives him an unusually clear inside view: he has seen frontier model development from multiple angles and has spent years translating complex AI concepts into accessible teaching material that has educated a huge portion of today’s AI engineers.

Part of a Bigger Picture

Karpathy’s take complements other high-signal perspectives on AI:

Jensen Huang (NVIDIA) views it primarily through infrastructure and hardware.
Marc Andreessen (a16z) sees it from the venture capital and startup ecosystem angle.
Larry Fink (BlackRock) focuses on capital allocation, markets, and economic implications.

Karpathy’s lens is the most hands-on and engineering-oriented. It’s also one of the most sobering and practically useful for anyone actually building or working with these systems day to day.

Also read:

The Rise of “Neuro-Slop”: American Companies Are Flooding Documents with AI’s Favorite Corporate Phrase

Death Tech in China: Family Creates AI Clone of Deceased Son to Protect Elderly Mother with Heart Condition

The $500,000 Corporate Retreat from Hell: Plex’s Survivor Trip in Honduras Nearly Killed the Team

AI Office Warfare in China: Employees Train AI to Replace Colleagues — Then Fight Back with Sabotage Tools

The Practical Takeaway

Prompt tricks and clever wording still have short-term utility, especially for quick experiments or when working with less capable models. But treating them as the core skill worth mastering long-term is missing the point.

The real skill — the one that will compound rather than decay — is learning to work with AI as a powerful but imperfect collaborator.

This means developing strong instincts about:

What tasks are worth delegating;
How to structure information so the model can do its best work (context engineering);
How to critique and refine outputs effectively;
When to trust the model and when to override it.

In short: the future belongs to people who bring deep understanding to the table and use AI to amplify it, not replace it. The models will keep getting better at the mechanical parts of thinking. Our enduring edge lies in knowing where we actually want to go — and recognizing when we’ve arrived at something worthwhile.

That distinction isn’t hype. It’s engineering reality, and it’s one of the clearest signals we have for how to navigate the next phase of AI productively.