@xiaohu: Claude Code's father's own CLAUDE.md is now just two lines... Claude Code team discusses "less is more" sharing how to communicate with models as capabilities increase: "Don't fight the model by adding more, because each generation of models gets stronger. What you painstakingly build today will soon be useless."

The CLAUDE.md of Claude Code’s creator? Just two lines for now…

The Claude Code team talks about “less is more” — how to communicate with models as they grow more capable:

“Don’t fight the model by adding more. Every generation gets smarter. Everything you painstakingly build today will be obsolete soon.”

Why does Claude Code insist on being a CLI instead of a GUI?

Because models improve too fast — a UI built today could be outdated in six months…

It comes down to four things:

1. Keep CLAUDE.md as short as possible. Regularly clear it out and start fresh.

His own CLAUDE.md is just two lines: auto-merge PRs, and send PRs to an approval channel. Everything else goes into the version-controlled, team-wide file they collaboratively update each week. When he sees teammates making avoidable mistakes, he just asks Claude on the PR to add a rule.

When the system prompt says “your CLAUDE.md is already several thousand tokens,” his advice is to delete it and rewrite: use the minimum to steer the model back on track. Add things back piece by piece only when the model goes off course. And you’ll notice — with each new model generation, you need to add less and less.

Many people’s problem is over-engineering.

2. Why insist on a CLI instead of a GUI?

Because models improve too fast to build a UI that won’t be obsolete in six months.

And a CLI actually lowers the barrier to entry — you don’t need to know Vim, Tmux, or SSH to use Claude Code. Just open it and it walks you through. There’s a Vim diehard on the team who says “you’ll have to pry Vim from my cold, dead hands,” but he himself uses VS Code and considers himself an ordinary engineer.

3. The tug-of-war between “verbose” and “concise” terminal output

Personally, he likes it verbose — he can glance at it and immediately see if the model has gone off the rails, then press Esc to stop it.

Half a year ago, he wanted to cut the verbose bash output, but the entire Anthropic staff revolted. Recently, they collapsed “read file / search file” into a one-line summary (which couldn’t have shipped six months ago because models frequently misread back then), and people on GitHub complained. So they added a verbose mode to satisfy both sides.

The refinement process: ship → use it yourself for a month → listen to users complain → iterate. He says his favorite thing is hearing how users actually want to use it.

4. Debugging with AI is already “absurd”

With good logging, you can just say “this object is broken,” and it will dig through logs, figure things out on its own, and even open a production channel to look at the live database.

The most striking example: he was tracking down a memory leak, did a heap dump, opened DevTools, combed through code for ages — nothing. His teammate Chris just threw the problem at Claude Code, which wrote its own little tool to analyze the heap dump and found the leak faster than he did.

Closing reflection

He says “what an agent can do” changes with every model generation. Newcomers often use it better than he, the veteran, does. “I have to keep re-adapting, because my brain is still stuck in the past.”

One sentence summary: Models are skyrocketing. The optimal human strategy isn’t to pile on configuration, scaffolding, and tools. It’s to subtract, stay light, cede judgment to the increasingly capable model, and constantly unlearn your own outdated habits.