Claude-Code

You open Claude, ask a question, and see an answer. That is the entire mental model most people have of a token bill. What is invisible on that same turn is that the session has already loaded a system prompt, whatever memory files apply, the tool definitions, the last few tool results, and — if you asked Claude to read a file — the file too. Add fifty turns and a couple of re-reads to fix a mistake, and the same conversation that felt like ten questions is closer to a hundred pages of input the model is re-processing every time you hit enter. ...

Three voices argue about how to extend a Claude agent — skills-first, MCP-anchored, complementary. The debate dissolves once you separate procedural knowledge from live reach.

A file-by-file walkthrough of the project-scoped Claude Code layout — CLAUDE.md, .mcp.json, settings, rules, commands, skills, subagents, and hooks — and the team-shared / personal-local split behind every file.

A concrete look at what turns a command-line tool into something an AI coding agent can drive reliably, with contrived before/after examples for structured output, subcommands, non-interactive defaults, exit codes, self-documenting help, idempotency, and when a token-oriented format like TOON is worth adding as a second output mode.

Claude Code is Anthropic’s AI-powered coding assistant — the one that reads files, runs commands, edits code, and drives a full development workflow from your terminal, editor, or a desktop app. It is a distinct product from the browser-based Claude chat at claude.ai; the two share the same underlying model but almost nothing about how you interact with them. Everything in this post is about Claude Code specifically. Out of the box, Claude Code behaves the same for everyone. But almost every serious user ends up with a personal configuration: a folder called ~/.claude/ on their machine that quietly shapes how Claude Code behaves for them. ...