Upgrade Your Claude Code Setup

CLAUDE.md is a plain-text file in your repo root (or ~/.claude/CLAUDE.md for global preferences) that Claude Code reads at the start of every session. It is the highest-leverage configuration change you can make. Put your stack, conventions, test commands, and critical constraints here — the things you would tell a senior engineer on day one. Claude Code follows these instructions in every tool call it makes in that repo. Global CLAUDE.md covers preferences that apply everywhere (verbosity, commit message format, no-test defaults). Repo CLAUDE.md covers project specifics: which framework, how to run tests, which files are off-limits, known gotchas. The combination eliminates most repeated corrections.

Model Context Protocol (MCP) servers extend Claude Code with new tools — think of them as plugins that give Claude Code access to external systems. The two highest-value servers to start with are: (1) Memory — lets Claude Code persist facts across sessions, so it remembers your preferences, project decisions, and recurring patterns without you repeating them. (2) Filesystem — gives Claude Code explicit read/write access to directories you specify, useful for multi-repo work. Install via: claude mcp add @modelcontextprotocol/server-memory and claude mcp add @modelcontextprotocol/server-filesystem /path/to/dir. Both are official Anthropic MCP servers. Verify installation with claude mcp list.

MCP servers exist for most major developer tools — GitHub, Supabase, Linear, Slack, Sentry, Cloudflare, and more. Once connected, Claude Code can query your GitHub issues, write Supabase migrations, update Linear tickets, and post Slack messages without leaving the terminal. High-value servers: @modelcontextprotocol/server-github (requires GITHUB_TOKEN), Supabase MCP (connects to your project via service role key), Linear MCP (create/update issues). Add them via claude mcp add with the appropriate environment variables. After connecting, test by asking Claude Code to 'list open GitHub issues' or 'show the last 5 Supabase migrations'.

Hooks let you run shell commands automatically before or after Claude Code tool calls — without interrupting the session. The two most useful hook events are PreToolUse (runs before Claude Code touches a file) and PostToolUse (runs after a file changes). Use cases: auto-run your linter after every file edit, run type checks before a commit, send a Slack notification when a long task finishes, log every tool call for auditing. Configure hooks in ~/.claude/settings.json under the hooks key. Each hook specifies an event, an optional tool name matcher, and a command to run. Hooks that exit non-zero block the action — use this as a gate for destructive operations.

Slash commands are reusable prompt templates stored as Markdown files in ~/.claude/commands/ (global) or .claude/commands/ (project-level). You invoke them with /command-name inside any Claude Code session. High-value commands to build: /review (code review checklist for your team's standards), /test (add tests for the current file following your test patterns), /deploy (run your deployment checklist step by step), /perf (profile and suggest optimizations). Each command file is a Markdown file whose content becomes the prompt. Use $ARGUMENTS as a placeholder for dynamic input. Example: /fix-ts-error runs a standardized TypeScript debugging flow with your project's specific patterns already included.

Extended thinking gives Claude Code significantly more compute for complex reasoning tasks — architecture decisions, debugging hard root causes, designing systems with many interacting constraints. Toggle it with the /think or /think-harder commands during a session. Best used for: designing a new database schema where tradeoffs matter, debugging a race condition across multiple services, writing a test strategy for a complex feature, evaluating refactoring approaches. Extended thinking costs more tokens and takes longer, so reserve it for problems where the higher quality output justifies the cost. Combine with sub-agents (using the Agent tool) for tasks that can be parallelized across multiple research threads.