Claude Code practical tips from Anthropic creator (Boris)

My notes

Summary

Boris, creator of Claude Code at Anthropic, gives a practical walkthrough of how to actually get value from the agentic coding tool. Key reveal: Anthropic technical onboarding dropped from 2-3 weeks to 2-3 days because new hires use Claude Code for codebase Q&A on day one. About 80% of technical staff at Anthropic use Claude Code daily.

Key Insight

• Start with codebase Q&A, not editing. First-day Anthropic hires do NOT start by writing code with Claude. They ask questions about the codebase. This teaches the boundary of what Claude can one-shot vs. what needs hand-holding.
• No indexing, no cloud upload. Claude Code does NOT index your code remotely. No setup wait. Code stays local. Models are not trained on your code.
• Onboarding metric: 2-3 weeks to 2-3 days for technical onboarding at Anthropic.
• Adoption metric: ~80% of technical staff at Anthropic use Claude Code daily.
• Plan mode trick: Don’t ask for 3000-line features one-shot. Just say “before you write code, make a plan, run it by me.” No special flags needed, the model knows what to do.
• Magic incantation: commit, push, PR - Claude infers commit format from git history, makes branch, pushes, opens PR. Nothing system-prompted; the model just figures it out.
• Iteration loop is the secret sauce. If Claude has a way to check its work (unit tests, Puppeteer screenshots, iOS simulator), it self-corrects. 2-3 iterations gets near-perfect results vs. one-shot misses.
• CLAUDE.md hierarchy: project root (checked in, shared), CLAUDE.local.md (personal, gitignored), nested directories (loaded on-demand), enterprise root (cross-repo policy). Keep it short, long files waste context.
• Permissions can be enterprise-policy enforced. Bash commands and URL allow/blocklists in enterprise config = employees can’t override. Same for MCP servers via checked-in .mcp.json.
• Hidden key bindings:
  • Shift+Tab to auto-accept edits mode (bash still prompts)
  • # to tell Claude to remember something into CLAUDE.md
  • ! to drop to bash mode, output goes into context
  • @ to mention files/folders
  • Escape to safely interrupt Claude mid-action
  • Escape twice to jump back in history
  • Ctrl+R to see full output Claude saw
• Multimodal from day one. Drag-drop image, paste image, file path, all work. Common pattern: drag-drop UI mock + tell Claude to implement against tier server, iterate.
• Why CLI not IDE? Two reasons, Anthropic devs use too many different IDEs to standardize, AND Boris bets that “by the end of the year, people aren’t using IDEs anymore” given model progress speed. Avoiding over-investment in UI.
• SDK = Unix philosophy. claude -p "prompt" with --allowed-tools, --output-format json. Pipe git status | claude -p or cat sentry.log | claude -p. “Super-intelligent Unix utility.”
• Power user pattern: parallel Claude sessions via tmux, SSH tunnels, git worktrees for isolation in same repo.
• Bash safety architecture: read-only command detection, static analysis to identify safe combinations, multi-level allow/blocklists. Built because manually approving every bash command kills productivity.