Boris Cherny, Creator of Claude Code, Shares 13 Pro Tips for Efficient AI Programming
Boris Cherny, the mind behind Claude Code, reveals 13 practical strategies to boost AI programming productivity—including parallel instance management...
Boris Cherny, creator of Claude Code, recently opened up about his personal workflow and 13 actionable tips to help developers use the AI programming tool more efficiently. He emphasizes that there’s no single "right" way to use Claude Code; its design supports both out-of-the-box functionality and flexible customization, with each team member adopting unique approaches.
Surprisingly, Cherny’s own setup is minimal. He notes Claude Code works exceptionally well without extensive tweaks, so he avoids over-customization. However, he has refined key strategies to enhance productivity:
First, Cherny runs multiple parallel instances: 5 labeled terminal sessions (1-5) with system notifications to alert which needs input, plus 5-10 web instances at claude.ai/code alongside local ones. He switches between local and web sessions using the teleport parameter and even starts sessions on his iOS app to check progress later.
For all tasks, he uses the Opus 4.5 model—calling it the best programming model he’s ever used. Though larger and slightly slower than the Sonnet version, its minimal need for human guidance and strong tool-calling capabilities make it more efficient overall.
Team collaboration is central: Cherny’s team maintains a shared CLAUDE.md file in git, updated weekly to document and fix Claude’s operational deviations. They use a Claude Code GitHub Action to tag @.claude in pull requests (PRs), adding relevant content to CLAUDE.md as part of the PR process—aligning with "Compounding Engineering" proposed by danshipper.
Workflow optimization includes starting sessions in Plan mode (double shift+tab) to refine PR plans before switching to auto-accept edits. Repetitive tasks are handled via slash commands (stored in .claude/commands/) like /commit-push-pr, which uses inline bash to precompute git status and save time. Sub-agents like code-simplifier and verify-app automate common post-task workflows.
Other tips: PostToolUse hooks to fix 10% of code formatting issues Claude misses; pre-authorizing safe bash commands via /permissions instead of skipping permissions; and integrating tools like Slack (via MCP server), BigQuery, and Sentry.
For long tasks, Cherny uses background validation, Stop hooks, or the ralph-wiggum plugin, running them in a sandbox to avoid permission interruptions. Most crucially, he stresses building validation mechanisms—like Chrome extension testing for UI changes—which can boost result quality by 2-3 times.
Sources