90% of Programming Skills Just Got Commoditized. The Other 10% Is Worth 1000X More.
There's a new layer between you and your code. It has no manual.
Andrej Karpathy recently wrote something that’s been rattling around my head:
“I’ve never felt this much behind as a programmer. The profession is being dramatically refactored as the bits contributed by the programmer are increasingly sparse and between. I have a sense that I could be 10X more powerful if I just properly string together what has become available over the last year and a failure to claim the boost feels decidedly like skill issue.”
He then listed what this new layer looks like: agents, subagents, prompts, contexts, memory, modes, permissions, tools, plugins, skills, hooks, MCP, LSP, slash commands, workflows, IDE integrations.
His conclusion: “Clearly some powerful alien tool was handed around except it comes with no manual and everyone has to figure out how to hold it and operate it, while the resulting magnitude 9 earthquake is rocking the profession.”
I felt this in my bones.
The Old Stack vs The New Stack
The old programming stack was hard enough:
Hardware -> OS -> Language -> Frameworks -> Your Code
Years of learning. Layers of abstraction. But at least it was deterministic. At least there were manuals. At least Stack Overflow had answers.
The new stack adds a layer on top:
You -> Prompts/Agents/Context/Memory/Tools/Modes -> Code
This layer is fundamentally different. It’s stochastic. It’s fallible. It’s unintelligible. And it changes every few weeks.
There’s no certification. There’s no textbook. There’s no “Effective AI Orchestration” by Joshua Bloch. Just a bunch of people figuring it out in Discord servers and sharing CLAUDE.md files and best practices like trading cards.
The Divide Is Already Here
Someone on Reddit described the pattern they’re seeing on their team:
“Two developers with similar experience working on similar tasks, but one consistently ships features in hours while the other is still debugging. At first I thought it was just luck or skill differences. Then I realized what was actually happening — it’s their instruction library.”
They’re watching an underground collection of power users share workflows like secrets:
Commands that automatically debug entire codebases
CLAUDE.md files that turn Claude into domain experts
Slash commands that turn 45-minute processes into 2-minute ones
Meanwhile, most people are still typing “help me fix this bug” and wondering why their results suck.
As one developer put it: “The differences between someone who opens up CC for the first time and someone with tuned md files is beyond night and day.”
The Skill Issue Is Real (But Not The One You Think)
Here’s what hit me about Karpathy’s framing: he called it a “skill issue.”
Not a tools issue. Not an access issue. Not a funding issue.
A skill issue.
The 10X boost exists. The leverage is real. But claiming it requires mastering something that didn’t exist two years ago and has no curriculum.
Someone in that same thread nailed the uncomfortable truth: “90% of traditional programming skills are becoming commoditized while the remaining 10% becomes worth 1000x more. That 10% isn’t coding — it’s knowing how to architect AI workflows.”
The irony is brutal. We spent years mastering syntax, frameworks, design patterns. Now an AI can generate all of that in seconds. What it can’t do is orchestrate itself effectively. That’s your job now.
What The New Layer Actually Looks Like
Let me make this concrete. Here’s what I’ve had to learn in the past year that wasn’t part of any CS curriculum:
CLAUDE.md Architecture
Your instructions file isn’t documentation. It’s programming. The structure, the phrasing, what you include vs exclude — these decisions compound across every interaction. A well-architected CLAUDE.md is worth more than a well-architected codebase.
Context Management
Every token matters. MCP servers eat context. Long conversations drift. You need to think about what Claude knows, what it’s forgotten, when to compact, when to start fresh. It’s memory management, but for a mind that isn’t yours.
Prompt Design
Not “prompt engineering” in the LinkedIn-influencer sense. Actual design. When do you give examples? When do you constrain? When do you let it explore? How do you phrase things so it doesn’t hallucinate? How do you trigger deeper thinking? These are learnable skills with massive payoff differences.
Tool Orchestration
MCP, skills, hooks, slash commands. Which tool for which job? When does an MCP server make sense vs a bash script vs a skill file? How do you chain them? How do you debug when the chain breaks?
Mode Awareness
Plan mode vs implement mode. When to let Claude explore vs when to constrain. When to use subagents. When to go linear. The meta of working with AI — knowing when to switch approaches — is itself a skill.
Verification Choreography
AI generates fast. Verification is the bottleneck. How do you structure your workflow so you’re not just rubber-stamping garbage? How do you catch the 8 production bombs before they ship? (Yes, I wrote about this.)
The Manual That Doesn’t Exist
An older developer on Reddit captured the frustration:
“I started using AI about 2 years ago. I thought I was doing good, but then I started seeing all this stuff about MCP servers, md files etc and I am kind of lost. I want to learn more and I want to improve my AI skills but it’s difficult for me.”
This is someone with decades of experience, feeling lost because the new layer has no onramp.
The manual doesn’t exist because the platform keeps shifting. Claude Code ships updates weekly. New features appear. Old patterns stop working. The MCP ecosystem is exploding. Skills just launched. Hooks changed. The ground won’t stop moving.
You can’t study for an earthquake. You can only practice surfing.
How I’m Learning (Imperfectly)
I don’t have this figured out. Nobody does. But here’s what’s working:
Steal shamelessly. Find people who are clearly more productive and reverse-engineer their setup. Their CLAUDE.md files, their slash commands, their workflows. GitHub repos, Discord servers, Reddit threads. The good stuff is scattered but findable.
Treat your setup as code. Version control your CLAUDE.md. Iterate on your slash commands. When something works, document why. When something fails, autopsy it. Your instruction library is a codebase now.
Invest in meta-skills. The specific tools will change. MCP might get replaced. Claude Code might get competition. But the meta-skills — context management, prompt design, verification choreography — those transfer.
Actually use the new features. Hooks exist. Subagents exist. Skills exist. Most people ignore them because they’re “advanced.” They’re not advanced. They’re just new. The learning curve is the moat.
Teach to learn. Writing about this forces me to understand it. Explaining my setup to others reveals the gaps. The best way to master the new layer is to articulate it.
The Uncomfortable Conclusion
Karpathy is right. There’s a 10X boost available. Failing to claim it is a skill issue.
But it’s a new skill. One that didn’t exist before. One that has no manual, no certification, no clear path.
The people figuring it out are building compound advantages. Every custom command, every refined CLAUDE.md pattern, every workflow optimization — it all stacks. The gap between those who master the new layer and those who don’t is widening fast.
The earthquake is still happening. The alien tool is still being figured out. The manual is being written in real-time by the people using it. And the rules are being changed as we speak/code/write.
Roll up your sleeves.
This is a companion to my previous essay on what Claude can’t do for you. That one covered the old skills that still matter. This one covers the new skills you need to add.