AI Coding Tools

AI coding tools split into two genuinely different interaction models rather than one uniform category: terminal-based agents like Claude Code and Codex CLI that live in a REPL and act on your real filesystem and shell, and editor-integrated tools like GitHub Copilot and Windsurf that layer suggestion-and-accept mechanics directly into a code buffer, with ChatGPT Desktop covering the more general conversational case accessible from anywhere on your system. Keyboard fluency matters more here than in most software categories, not less, precisely because these tools are structured around a rapid interrupt-accept-reject loop rather than a fixed sequence of menu actions — a terminal agent that's about to run the wrong command needs an instant Ctrl+C, not a hunt through a menu, and an inline AI suggestion that's mostly right but not quite needs a fast partial-accept rather than an all-or-nothing choice. The permission and approval mechanics distinguishing these tools from ordinary software are also genuinely consequential rather than cosmetic: Claude Code's Shift+Tab mode cycling, Codex CLI's approval-mode configuration, and Windsurf's per-step Cascade review all exist because these tools can take real actions — writing files, running commands — that a plain autocomplete or chatbot never could, and knowing the exact keystroke that stops or reviews that action before it happens is a meaningfully higher-stakes piece of keyboard knowledge than memorizing a formatting shortcut in a word processor.

Developers moving between a terminal agent and an editor-integrated assistant in the same session benefit from understanding that these tools deliberately borrow different amounts from traditional shell conventions versus GUI editor conventions — recognizing which model a given tool follows is often the fastest way to guess correctly at an unfamiliar binding before ever needing to look it up.