Ex-Tesla AI head has seen a 'phase shift in software engineering' using Claude Code — and his manual skills slowly 'atrophy'
Michael Macor/The San Francisco Chronicle via Getty Images
- Andrej Karpathy posted his "notes from Claude Coding," describing a recent "phase shift" in software engineering.
- "I really am mostly programming in English now," Karpathy wrote, describing the slow "atrophy" of his manual coding skills.
- Engineers from xAI and Anthropic chimed in with their own takes on agentic coding.
He coined "vibe coding." Now, he sees a "phase shift" in software engineering.
Andrej Karpathy is one of AI's guiding figures. He was a founding member of OpenAI and later served as Tesla's director of AI. He also coined the term "vibe coding," the AI-assisted coding movement that has taken software engineering by storm and was named Collins Dictionary's word of the year.
In his "random notes from Claude Coding" — which are over 1,000 words long — Karpathy wrote about the changes to his own coding style. Posted on X on Monday, the notes have already elicited reactions from engineers at Anthropic, xAI, and more.
AI coding agents "crossed some kind of threshold of coherence around December 2025 and caused a phase shift in software engineering," Karpathy wrote.
A few random notes from claude coding quite a bit last few weeks.
— Andrej Karpathy (@karpathy) January 26, 2026
Coding workflow. Given the latest lift in LLM coding capability, like many others I rapidly went from about 80% manual+autocomplete coding and 20% agents in November to 80% agent coding and 20% edits+touchups in…
Karpathy name-dropped both Anthropic's Claude Code and OpenAI's Codex as having significant improvements. Claude Opus 4.5, the model that has garnered much love from engineers online, came out at the tail end of November.
The AI leader's workflow has changed as a result of the AI tools. From November to December, Karpathy's 80/20 ratio flipped. He once used 80% manual coding and 20% agents; now, it's 80% agents and 20% manual code editing.
"I really am mostly programming in English now, a bit sheepishly telling the LLM what code to write... in words," he wrote.
The change to AI-written code "hurts the ego," but is too powerful to ignore, Karpathy wrote. He also devoted a whole section of his notes to the "fun" he has while coding with large language models.
What of those traditional coding skills, the ones you learn in a computer science program or through endless digital courses? That's a whole other function, Karpathy wrote, and one that might decline.
"I've already noticed that I am slowly starting to atrophy my ability to write code manually," he wrote.
In Karpathy's comments, engineers from leading AI companies sounded off. Ethan He, an xAI engineer and Nvidia alum, wrote that a "10x engineer can be a one-man army."
Charles Weill, another xAI engineer, wrote that founders can now "divide themselves" with coding agents, like a VC divides their capital over a portfolio of companies.
Boris Cherny, an Anthropic staffer and the creator of Claude Code, wrote that he read Karpathy's "thoughtful" post till its end.
The Claude Code team at Anthropic may offer a model of where the industry is moving, Cherny wrote. His team is "mostly generalists" and filled with 10x engineers.
"Pretty much 100% of our code is written by Claude Code," Cherny wrote. "For me personally it has been 100% for two+ months now, I don't even make small edits by hand."
The Anthropic employee also acknowledged the "quality" problems with AI-written code. Agents can overcomplicate things and can leave around dead code, he wrote.
His solution: having AI review the AI-written code.