A reflective piece arguing that the proliferation of AI tools is turning professionals into prompt-dependent 'managers' rather than true craftsmen, and warning that those who maintain their own foundational skills will command higher value.
Look around. Every other product launching right now is some variation of "AI-Powered \[insert buzzword\]." They're everywhere. Modern tools have given founders and developers a convincing illusion of omnipotence: idea hits, feed it to an LLM, stack some agents on top, and MVP is done in a weekend. https://preview.redd.it/37ocn6azkv1h1.png?width=1672&format=png&auto=webp&s=06d4a9ef986d56a9eb3417e67a3524c18e73e100 Sounds great, right? On the surface, yes. But underneath that fast-launch facade, something is quietly rotting: thinking is getting commoditized, and we're losing craft. Real mastery in any field takes years of practice, failure, and deep focus. Today, apparently everyone is a master for $20 a month. That's a lie we're telling ourselves. Just look at how much panic a 5-hour rate limit window in Claude generates online. Tokens run out, and suddenly people have two options: wait for the reset like a metered parking spot, or upgrade. It's like a Michelin-starred chef who can no longer taste food, just dictating to a chatbot: "make me a pasta." Without the subscription, he can't cook. **The counterargument: "But orchestrating AI IS the new skill."** Fair. But it's a horizontal skill, not a vertical one. You learn to coordinate agents while losing deep domain knowledge. Think conductor versus virtuoso violinist. A conductor is impressive - but if the orchestra walks off stage, can he play a solo that makes the room go quiet? This is most visible in developers right now. People who got used to copy-pasting from Cursor or Claude hit a wall on hard architectural problems. When a product grows, starts needing real trade-offs, starts buckling under load - prompts stop working. The muscle for hard problems atrophied because they never had to build it. Same thing is happening to analysts, marketers, designers, researchers. # My position: barbell, not crutch Running out of tokens doesn't scare me. My foundation means I can work regardless of what's left in my quota, whether there's internet, whether a subscription is active. The only thing that throws me off is running out of good coffee. I use LLMs heavily. But with one condition: AI is a barbell, not a crutch. It sharpens my own work - it doesn't replace the parts I care about. The fastest, most tireless junior I've ever hired. But the senior judgment and the final call always stay with me. # Two types of professionals The market is already splitting into two groups. **Token-dependent:** live limit to limit, panic when Anthropic or OpenAI have an outage, can't produce anything original without a prompt to lean on. **Token-independent:** use AI as a force multiplier but can, at any moment, sit down and do the work themselves - with more depth, more precision, better judgment. The second group will command much higher rates. When the world is drowning in mediocre AI-powered software and content - and it will be - clients and employers will pay serious money for people who actually understand what they're building and why. Curious whether others are feeling this shift. Are you building toward token-independence, or does the dependency not bother you?
The article argues that AI agents should not just obediently execute tasks but should proactively challenge humans when tasks are vague, contradictory, or risky, transforming from tools into true collaborators.
The article critiques the proliferation of AI-generated work in the workplace, where employees use tools like Claude to produce expert-seeming outputs without genuine expertise, leading to systemic issues in management and accountability.
The article observes a trend where junior AI engineers focus on high-level tools like prompt engineering and low-code platforms rather than deep understanding of fundamentals, raising concerns about problem-solving skills in interviews.
Andrew Ng predicts traditional prompt engineering will become obsolete within six months, shifting focus toward autonomous agentic systems that iterate and self-improve with minimal human input.