A contributor submitted a pull request to Tailwind's GitHub on November 18, 2025.
The PR was simple. Add an endpoint called `/llms.txt` that would serve a text-only version of all Tailwind documentation, optimized for large language models to consume. Extract text from MDX files. Remove JSX components. Preserve code blocks. Make it easier for Claude and ChatGPT to read the docs.
Nobody responded.
For seven weeks, the PR sat open. Then on January 6, someone pinged the thread. "Why is this one not moving?"
Adam Wathan, Tailwind's founder, finally replied: "Have more important things to do like figure out how to make enough money for the business to be sustainable right now."
The next sentence explained everything.
"75% of the people on our engineering team lost their jobs here yesterday because of the brutal impact AI has had on our business."
Three engineers. Gone. Revenue down 80%. Documentation traffic down 40% since early 2023. Not because Tailwind isn't popular, it's bigger than ever. But because developers stopped visiting the documentation where they'd discover the paid products.
They're asking Claude Code instead.
Here's what twelve months changed.
Claude Opus 4.5 scored higher on performance engineering exams than any human candidate ever. It autonomously refined its own capabilities in four iterations, peak performance that other models couldn't match after ten. Claude Code worked for an hour straight creating hundreds of files, deploying a working website without human input.
OpenAI's Codex went from autocomplete to agentic engineer. GPT-5.2-Codex runs 30-minute autonomous coding sessions in large repositories. It completes complex refactors and code migrations without losing track. A security researcher used it to find and disclose vulnerabilities in React that required sustained, multi-step reasoning.
Both tools can now generate complete Tailwind implementations,spacing, shadows, that perfect shade of gray, in minutes. The $299 UI kit with fifty carefully crafted components? Five minutes. Maybe less.
The irony landed on GitHub: Someone wanted to make Tailwind documentation *more* accessible to LLMs. But LLMs were already killing the business. Traffic to docs fell 40% because developers were using AI tools that generated code without ever visiting the site. The PR that promised to help was exactly the problem.
DeepSeek's R1 model cost $5.6 million to train. American equivalents spent hundreds of millions. The capability gap closed overnight. The price collapsed. Then the models kept improving.
What Tailwind sold was never really the code. It was the shortcut. The trusted answer. The decision you didn't have to make about spacing for the hundredth time.
Now Claude Code makes those decisions. Codex makes those decisions. In terminal windows. In IDEs. In cloud sandboxes that spin up in seconds. For twelve dollars a month.
The documentation that used to teach people now teaches models. The components that used to convert customers now generate on demand. The business model that funded one of the web's most popular frameworks just said goodbye to three quarters of its team.
This was January 6, 2026. The layoff day.
The PR closed the same day. Still no merge.
Between last January and this one, AI didn't just get better at coding. It got better at being the shortcut. At replacing the moment when you'd search for an answer and land on someone's documentation. At making the visit, and the discovery that follows, obsolete.
What do you sell when the thing you packaged becomes the thing AI generates as a side effect?