Boris Cherny built Claude Code at Anthropic. The former Meta principal engineer now heads the project and hasn’t written production code by hand since late 2025. He ships dozens of pull requests a day, often from his phone, and on some days orchestrates tens of thousands of AI agents at once.
That shift started with a terminal tool. Claude Code turns the model into a persistent agent that reads files, runs commands, edits code, and spawns sub-agents inside your repo. Installation remains straightforward for most developers.
Run the native installer on macOS or Linux with curl -fsSL https://claude.ai/install.sh | bash. Windows users use the equivalent PowerShell command. Then authenticate with claude login or claude auth login. Launch a session in any project directory simply by typing claude.
Power users immediately run /init to generate a starter .claude/CLAUDE.md file. They also check /permissions, /allowed-tools, and /memory to tune behavior from the first session. Dark theme and notifications come via /config or /theme.
Context engineering separates average sessions from production-grade output.
Claude Code has no built-in long-term memory across restarts. The solution lives in markdown files that load automatically. Place a project-level .claude/CLAUDE.md in the repo root. Add a global version at ~/.claude/CLAUDE.md for personal standards. Use .claude/local.md for session-only overrides.
A strong CLAUDE.md file lists the tech stack, coding principles, forbidden patterns, naming rules, and architecture decisions. One common template starts with headings for Tech Stack, Coding Principles, Forbidden Patterns, Naming Conventions, and Architecture Decisions. Keep the file under roughly 200 lines so the model actually reads it. Anthropic’s own best-practices guidance stresses this file as the primary way teams share state across sessions.
Auto-memory features let Claude append high-signal lessons during work. Enable it through the memory editor. Review and prune regularly to avoid bloat. Commands like /compact and /btw help keep context tight during long runs.
Three core workflows dominate daily use. Start every new codebase or large task with targeted questions: “Trace the authentication flow end-to-end” or “Explain changes in PR #1243.” Claude Code answers with file paths and line numbers after searching the repo and git history locally.
The plan-execute-review loop prevents wasted edits. Issue /plan "Implement user profile editing with optimistic updates", review the output, then approve or refine. After changes, run /diff followed by /code-review or the ultra variant for deeper analysis. Many teams add /code-review --fix to apply fixes automatically.
Sub-agents scale output further. The /agents command spawns specialized instances—one for tests, another for documentation, a third for refactoring. Worktrees via /batch keep large refactors isolated and safe. Background mode with /background detaches long tasks so the main session stays responsive.
Slash commands form the daily vocabulary. /plan enters planning mode. /cd changes directories while preserving cache. /resume and /branch fork conversations. /doctor troubleshoots sessions. /rewind rolls back edits. Custom commands live in .claude/commands/. Skills and hooks package repeatable patterns; load a shared skill with /skills boris or similar.
Model Context Protocol servers extend reach without custom glue code. Connect to GitHub, internal databases, CI logs, or monitoring tools through MCP. Recent updates include tool search that dynamically loads only needed descriptions, cutting context usage dramatically when many servers are active. Configure via .mcp.json or the global equivalent. Claude Code treats MCP tools as native callables once registered.
Prompt templates that work consistently start with project memory, then state the task, constraints, success criteria, relevant files, and desired output format. A debugging template includes the exact error, asks the agent to reproduce the failure, fix it, and run the affected test suite. Refactor prompts specify the target pattern, require 100 percent behavior preservation, and demand added tests plus documentation updates.
Integration points multiply leverage. The same engine powers terminal, VS Code, JetBrains, desktop, and web surfaces. A GitHub App lets you tag issues or PRs for automatic Claude responses. Long-lived tokens support CI/CD pipelines. IDEs show no lock-in—switch surfaces without losing memory files or MCP setups.
Real usage data shows clear patterns. Individual engineers report 50 to 90 percent of their code now originates inside Claude Code sessions. Teams inside Anthropic sustained productivity gains even while headcount grew rapidly. Onboarding time for new engineers drops from weeks to days once strong CLAUDE.md files exist. Large refactors finish in hours rather than days. Test coverage rises because writing and running tests becomes trivial.
Common failure modes trace back to weak context, skipped reviews, or vague instructions. Fix the first by investing time in memory files. Fix the second by always running code review plus manual verification. Fix the third by naming exact files, examples, and measurable outcomes in every prompt.
Advanced configurations mirror Cherny’s own setup.
Power users maintain a global CLAUDE.md with personal philosophy, project-specific memories, a library of custom skills, background agents for linting and dependency updates, and multi-agent review pipelines before any merge. Some run scheduled cloud routines that email daily PR summaries or security scans. Voice input pairs naturally with screenshot analysis in desktop mode for visual debugging.
The 27-minute workshop video referenced across recent X posts and YouTube remains a strong entry point. Search for “Mastering Claude Code in 30 Minutes” featuring Boris Cherny directly. It covers installation, memory files, planning loops, and basic prompting. Treat it as the starting point, not the destination. Daily deliberate practice with the commands and memory structures above compounds faster than any single video.
Recent coverage in Fortune (June 8, 2026) notes Cherny managing fleets of agents while shipping from his phone. Lenny’s Podcast episode from February 19, 2026, captures his view that coding for many routine tasks is effectively solved, shifting focus to harness design and agent orchestration. Anthropic’s own engineering blog posts detail MCP improvements and best-practice patterns for parallel sessions.
Developers who treat Claude Code as a full teammate rather than autocomplete see the largest gains. They invest in context files, design explicit agent loops, review output rigorously, and keep architectural decisions in human hands. The tool supplies the execution layer. The engineer supplies direction and standards.
Start in your primary repo today. Run the installer, authenticate, execute /init, and spend thirty minutes on a solid CLAUDE.md. Then ask real questions about your own codebase. The productivity difference appears within the first few sessions.
WebProNews is an iEntry Publication