What I Learned from How Anthropic's Team Actually Uses Claude Code

ยท 2 min read

Boris , who works at with Anthropic, recently shared his Claude Code setup. What struck me wasn't any single tip but the philosophy behind it: "surprisingly vanilla, customizable & hackable, no one correct way."

As someone who works with and tests agents for a living, I've been stress-testing Claude Code for real enterprise use-case workflows. Here's Boris's framework with my own takeaways.

Parallelization is the unlock

Boris runs 5-10 Claude sessions simultaneously (expensive), some in terminal, some on claude.ai/code, handing off between sessions them with teleport.

This mirrors how I one can use these tools in the most effective way: the bottleneck isn't the model, it's your ability to keep multiple work-streams moving.

If you're still using Claude Code daily, you're leaving speed on the table.

Opus 4.5 with thinking, always

His default is Opus 4.5 with extended thinking enabled. Counterintuitive because it's slower per-response, but Boris argues it's faster end-to-end: less steering, better tool use, fewer corrections.

I've found the same that the "fast" models often costs more in back-and-forth. In cases that the output of fast, or cheaper model is subpar, I end up spending more time analyzing the contents for accuracy, and the back-and-forth iterations for refinement ends up taking more time, and tokens. With thinking, CC is likely to give an output that is more aligned, and coherent to the goal at hand.

Plan mode before auto-accept

This one changed my workflow. Start sessions in Plan mode (shift+tab twice), write out your PR goal, iterate on the plan with Claude, then switch to auto-accept for execution.

The insight: a good plan compresses implementation time dramatically.

CLAUDE.md as institutional memory

Boris's team shares a single CLAUDE.md checked into git. Multiple contributors update it weekly. They even tag @claude on PRs to capture learnings automatically via GitHub Actions.

This is compounding engineering โ€” every session makes the next one smarter. I'm experimenting with something similar for my own projects.

The meta-lesson

The cheatsheet ends with: "There is no wrong way, just your way. Enjoy building with Claude Code."

Most people are hunting for the "optimal" setup. Boris's point is that the optimal setup is the one you'll actually iterate on and the one that actually works for YOU.

Start from small steps. See how they improve your workflows better. Iterate, refine and tune it. Make your vibe coding set-up truly yours and results will follow.