Claude Code Setup

This is my personal Claude Code config. It sets up global defaults, hooks, skills, and MCP servers that I’ve found make a meaningful difference in day-to-day use.

Run this on any machine that has Claude Code installed:

curl -fsSL https://nadia.fyi/claude/install.sh -o /tmp/claude-install.sh && bash /tmp/claude-install.sh

The script is interactive — it will prompt you to confirm each section (hooks, skills, peon-ping) before installing.

What gets installed:

~/.claude/settings.json — sets acceptEdits mode (no prompt on file edits), adds safe allow/deny rules, enables auto-compact, and wires in the hooks and MCP servers below.

~/.claude/CLAUDE.md — global instructions that apply to every project: proactively use subagents, consider multiple approaches before implementing, plan before coding, make surgical changes only.

Hooks — stop-notify.sh sends a macOS notification when Claude finishes a task. pre-tool-block.sh adds an extra safety layer blocking destructive shell patterns. post-edit-format.sh auto-runs prettier after file edits when a prettier config exists in the project. stop-autocommit.sh creates a WIP git checkpoint after each subagent completes.

Skills — /karpathy forces the agent to state its interpretation, justify its approach, constrain scope, and define done criteria before writing any code. /architect spawns parallel subagents to evaluate multiple technical approaches and compare them before implementation begins.

MCP servers — Context7 for real-time library documentation (eliminates stale-docs errors). Sequential Thinking for structured multi-step reasoning chains.