DHH, the creator of Ruby on Rails, has rewritten a Python library in Rust using only Anthropic's Claude. The project consumed 11 million tokens, ran for just over three hours, and reduced startup time from 87 milliseconds to 2 milliseconds - a 9.6x rendering speed increase. The result: a tool that once required a Python runtime is now a standalone 3MB executable. The post drew commentary from Anthropic's Claude Code team lead, who called the feat "absolutely amazing."
Just over a year ago, DHH was publicly opposed to handing his keyboard to AI. In a widely-read essay in May 2024, he argued that AI should remain an auxiliary "pair programmer" at most, and said: "If one day I have to completely hand over the keyboard to AI, I would rather retire." In a six-hour interview with Lex Fridman, he said every line of his code was typed by his own hands.
By April 2025, his position had shifted. During a podcast, he said he had fully moved to an "Agent-first" mode. Now, after rewriting the TerminalTextEffects library (TTE) with Claude, he has gone further, predicting a near future where coding as we know it changes shape.
How an 11-million-token rewrite worked
TTE - TerminalTextEffects - is a terminal visual effects engine that developers can import as a Python library or run as a command-line program. It controls text movement, color, gradient effects, and animation sequences, making terminal UIs more visually dynamic.
DHH says the rebuild involved eight agents working simultaneously on separate branches, merging their output in batches. The process ran for 3 hours and 11 minutes, producing 21,000 lines of Rust - nearly the same size as the original Python codebase. He asked Claude to draw up a plan, then intervened once. He later explained there was no complex workflow, no preset goals, and full delegation of authority.
When he ran the same task through OpenAI's Codex, the result was also solid but notably different: 30% slower and $43 in token costs.
What this means for how software gets built
DHH's experiment illustrates a broader shift. Jarred Sumner, founder of Bun, led his team to migrate the Bun codebase from Zig to Rust - a million-node project - in under two weeks. Years ago, a work of that scale would have required several quarters or more.
OpenAI researcher roon has been blunt about his own process: "I don't write code anymore, 100% of it is done by GPT." Claude Code's originator has claimed he does not code manually and can merge 27 pull requests in one day, all handled by an agent.
DHH's follow-up remarks have raised eyebrows among developers. When pressed on what Ruby's position would be in five years, he said that few people would be writing or reading code manually by then, and that the concern "really doesn't matter." That comment understands changing roles, not a disappearing profession.
The core work of programming shifts toward defining requirements, designing constraints around them, building test systems, and evaluating whether the AI's implementation is correct, rather than writing each line by hand. The moat becomes insight into what is worth building and judgment about whether it works.
Why this matters for AI for IT & Development
For software developers, the DL HH experiment shows what AI delegation actually looks like in production - with 21,000 lines of code, 11 million tokens, and a measurable performance improvement. The skills to focus on are the ones the passage demonstrates: writing clear task plans, verifying outputs, and reviewing tests, not typing each line. It's a concrete reason to workshop with AI-assisted workflows, not a theoretical roadmap for developers to put both greasy hands behind their heads.
Teams that treat AI as a junior engineer who executes defined tasks - and evaluate its output rigorously - will get the results DHH did. Teams that hand over vague requests and expect magic will not.
A AI for IT & Development approach that takes account of real-world agentic coding can help developers learn how to define and evaluate tasks for AI, rather than just generate code. That kind of planning will matter more as the marginal cost of generating code continues to fall.
For software developers making the AI Learning Path for Software Developers, the shift means learning to treat AI as a tool you direct - with review built-in - not a magic box that produces problems. The experiment suggests that "one-shot" prompts for a single well-scoped task work, and that the quality of your instructions matters more than the number of lines you type.
Your membership also unlocks:
