Keep pulling the thread on Boris Cherny.
Anthropic has released Claude Code, a command line tool for agentic coding.
Anthropic recommends using the --dangerously-skip-permissions flag for Claude Code only within a container that does not have internet access to minimize security risks.
Claude Code is designed to be a low-level and unopinionated tool that provides near-raw model access without imposing specific workflows.
Claude Code automatically pulls context from a special file named CLAUDE.md when a conversation starts.
By default, Claude Code requests user permission before performing actions that could modify the system, such as file writes or running certain bash commands.
Claude Code can use the gh command-line interface to interact with GitHub for tasks like creating issues and pull requests.
Claude Code can function as both a Multi-agent Controllable Peripherals (MCP) server and client.
Users can create custom slash commands in Claude Code by storing prompt templates as Markdown files in the .claude/commands folder.
The phrases "think", "think hard", "think harder", and "ultrathink" map to progressively larger computation budgets in Claude Code.
Instructing Claude Code to research and create a plan before writing code significantly improves its performance on complex problems.
Claude Code's performance is best when it is given a clear target to iterate against, such as a visual mock or a test case.
Claude Code has a --dangerously-skip-permissions flag that bypasses all permission checks for uninterrupted operation.