Boris Cherny, creator of Claude Code, shares ten usage tips drawn directly from the Claude Code team at Anthropic. He notes upfront that there is no single method and encourages experimentation.
Parallelism via git worktrees: The team's number-one tip is to use 3 to 5 git worktrees simultaneously, each running its own Claude session. This approach is identified as the biggest productivity gain. Some team members create shell aliases (za, zb, zc) to navigate quickly between worktrees, while others keep a dedicated worktree for log analysis and BigQuery.
Systematic plan mode: For any complex task, the team recommends starting in plan mode and investing the effort needed so Claude can implement it in a single pass. One practice is to use one Claude to draft the plan, then a second to review it with the eye of a "staff engineer." As soon as an implementation goes off track, immediately return to plan mode rather than pushing through.
CLAUDE.md as evolving memory: After every correction, ask Claude to update its CLAUDE.md file so the mistake isn't repeated. Claude turns out to be remarkably effective at writing its own rules. One engineer has Claude maintain a directory of notes per task/project, updated after every PR.
Skills and custom commands: Any action repeated more than once a day deserves to become a skill or slash command. Examples cited: /techdebt to identify duplicated code, a command that syncs 7 days of Slack, GDrive, Asana, and GitHub into a single context, or analytics-engineer-style agents for dbt.
Autonomous bug fixing: Enable the Slack MCP, paste a bug thread, and simply say "fix." Or ask "Go fix the failing CI tests" without micromanaging. Claude also handles troubleshooting via docker logs on distributed systems.
Advanced prompting techniques: Ask Claude to act as a reviewer ("Grill me on these changes"), to prove the code works via a diff between branches, or to redo a mediocre implementation with "implement the elegant solution."
Optimized terminal environment: The team favors Ghostty for its synchronized rendering and unicode support. They use /statusline to display context and the git branch, tabs colored by task, and voice dictation (fn x2 on macOS) for prompts 3x faster.
Subagents and analytics: Add "use subagents" to complex requests to delegate compute. The team uses Claude with the BigQuery CLI (bq) for all their analytics queries, and hasn't written SQL by hand in over six months.
Continuous learning: Enable the "Explanatory" or "Learning" style in /config, generate HTML presentations or ASCII diagrams to understand code, and create a spaced-repetition skill to consolidate learning.