The mechanics in one place — the agentic loop, slash commands, hooks, MCP servers, subagents, and Skills. Opens here on the page.
Claude Code is an agentic loop, not a chat box. You give it a goal; it reads files,
runs commands, edits code, checks the result, and repeats until the goal is met or it needs you.
Everything below is a way of shaping that loop.
The permission model
Every action that touches your machine — writing a file, running a command — is gated. You
approve once, always for that pattern, or not at all. This is the single most important thing to
understand: Claude Code cannot quietly do something you did not allow.
Slash commands
Typed at the prompt. A few worth knowing on day one:
/init — writes a first CLAUDE.md by reading the project/clear — starts a fresh context; the fix for a session that has wandered/config — model, theme, and session settings/help — the full list, which is longer than this
You can write your own: a Markdown file in .claude/commands/ becomes
/your-command. That is the whole mechanism.
CLAUDE.md — the memory that persists
A CLAUDE.md at the project root is read at the start of every session. A
CLAUDE.md in ~/.claude/ is read in every project. This is the
file the whole curriculum is about — Lessons 1 through 4 build it.
Subagents
A subagent is a separate Claude with its own context window, given one task, reporting back a
result. The point is context hygiene: exploring twenty files to answer one question fills a
context window with noise, so you send a subagent to do the reading and return the answer.
Define them in .claude/agents/.
Skills
A folder with a SKILL.md describing when to use it and how. Claude loads it only
when the task matches, so twenty skills cost nothing until one is relevant. Use them for
procedures you would otherwise re-explain every time.
Hooks
Shell commands the harness runs at fixed points — before a tool call, after one, when a session
ends. Configured in settings.json. Hooks are how you get behaviour that must happen
every time rather than when Claude remembers: run the formatter after every edit, block writes to
a protected path, log what changed.
MCP servers
The Model Context Protocol is how Claude Code reaches things that are not files — a database, a
ticket tracker, a design tool, an internal API. A server exposes tools; Claude calls them under
the same permission model as everything else. Lesson 8 covers this properly.
Written for this curriculum and kept alongside it, so nothing here sends you
off-site mid-lesson.
