⇠ chan.dev / posts

Claude Code - Patterns and Practices

🌱 This post is in the growth phase. It may still be useful as it grows up.

A collection of patterns, practices, and learnings for using Claude Code effectively.

2026

Boris Cherny’s Setup

Boris Cherny created Claude Code and shared how he personally uses it. Key insight: his setup is surprisingly vanilla - Claude Code works great out of the box.

How to work like Boris:

  • Run Claude in parallel: Keep 10-15 sessions going simultaneously - 5 in terminal with OS notifications, 5-10 in browser. Treat AI as schedulable capacity, not a single-use tool.
  • Default to Opus 4.5 with thinking: It’s slower, but requires less steering. A right slow answer beats a wrong fast answer when you factor in iteration cost.
  • Maintain shared CLAUDE.md in git: When Claude makes mistakes, document them immediately. Update it multiple times per week. Turn errors into team-wide guardrails.
  • Tag Claude in PR reviews: Set up a GitHub action so Claude can update CLAUDE.md based on code review feedback. Embed agents into your collaborative workflow.
  • Iterate in Plan mode first: Don’t jump straight to auto-accept. Work through the plan until you’re satisfied, then switch modes for implementation.
  • Build slash commands for repetitive work: Create custom commands like /commit-push-pr for workflows you run dozens of times daily.
  • Use subagents as workflow primitives: Treat specialized agents (code-simplifier, verify-app) as reusable building blocks, not monolithic solutions.
  • Automate formatting with PostToolUse hooks: Let hooks handle code formatting automatically to prevent CI failures.
  • Pre-allow safe commands via /permissions: Don’t skip permissions entirely - use /permissions to whitelist safe bash commands across your team.
  • Share MCP configs across the team: Check in configurations for tools like Slack, BigQuery, and Sentry so everyone benefits.
  • Always provide verification loops: Give Claude browser access to test UI changes iteratively. This 2-3x’s the quality of results.
  • Optimize for iteration cost, not token speed: You don’t trust AI - you instrument it. Build systems that reliably produce what you need.

Full thread:

Sources: