Most developers using Claude Code are leaving 2-3x output on the table. Boris Cherny - the engineer who built Claude Code - runs a very specific setup. Here's the full breakdown so you can copy it. First up: 🖥️ Session Management → Open claude.ai/code in Chrome alongside your terminal sessions → Hand off a terminal session to the browser using & in your terminal → Start sessions from your phone in the morning and pick them up on desktop later → Run multiple sessions in parallel without losing context → Never restart from scratch mid-task again → Keep long-running work alive across devices → Treat your browser as a full Claude Code workspace → Switch contexts without breaking your flow Next: 🧠 Model Selection → Use Opus 4.5 with thinking enabled for every task → It is slower per response but you steer it far less → It is better at using tools than lighter models → Net result: almost always faster overall → Stop babysitting the model through every step → Let it reason its way through complex tool use → Reserve fast models for truly trivial tasks only → Default to the smarter model and save time No more micromanaging every output. Your setup should feel like: "It just handled that on its own." Then: 📝 Memory and Commands → After every correction, end with: "Update your CLAUDE.md so you don't make that mistake again." → If you do something more than once a day, make it a slash command → Boris uses /commit-push-pr dozens of times every day → Commands live in .claude/commands/ and get checked into git → The whole team inherits your best workflows automatically → Run /install-github-action to enable PR automation → Switch to auto-accept edits mode for uninterrupted implementation → Shared commands eliminate inconsistency across the team Every correction you give becomes permanent institutional knowledge. Finally: ✅ Verification Loops → Add "Open a browser, test this end to end, and fix anything that does not work" to the end of any session → Claude tests using the Chrome extension - opens a browser, tests the UI, iterates until it works → A verification loop produces 2 to 3 times better output than no verification → Run code-simplifier to clean up after Claude finishes → Run verify-app to test end to end before handing back → Use Stop hooks or the ralph-wiggum plugin for long-running task verification → Never use --dangerously-skip-permissions - run /permissions and pre-allow specific safe commands → Check .claude/settings.json into git so the whole team shares the same safe defaults Boris calls the browser verification loop the single most important thing in his entire setup. Just the right model, persistent memory, and a verification loop that closes the gap between "it ran" and "it actually works." Comment AGENTS and I'll send you Boris Cherny's full Claude Code setup breakdown.