Claude Code Accelerator / Reference

How Claude Code Works

One loop sits underneath the features. Claude reads a window, calls tools, sees the results, and repeats. The rest of the setup controls what goes into that window, when it loads, and what Claude is allowed to do.

Every feature answers one of three questions

What is in the window
What Claude knows right now. It forgets everything between sessions.
When it loads
Right away at startup, or only when it is needed. The window is finite.
What Claude can do
Which tools it can reach for, and what is off limits.
Read top to bottom. The atom is the foundation. Each numbered rung builds on the one above it, and the grey ribbon names the problem it fixes. Need the folder map too? Open setup sketches.
The atom · everything rests on this

Claude Code is one loop

A window of text is assembled, Claude reads it and either replies or calls a tool, the tool output returns to the window, and the loop repeats until the work is done.

1. Assemble the window 2. Read it, reply or call a tool 3. Tool output returns to window 4. Repeat
It forgets between sessions. Anything you want it to know has to be loaded into the window again.
The window is finite. A full window is slower and dumber than an empty one. Attention rots as it fills.

The whole board in one sentence each

The loop
Claude reads a window and reaches for tools, and forgets it all between sessions.
CLAUDE.md
A note loaded into every window so it stops starting from zero.
Context management
The window is finite, so you manage the budget.
Skills
Cheap descriptions always loaded, full instructions loaded on demand.
Slash commands
Skills you fire by name when you want a guaranteed button.
MCP
New tools and data plugged in from outside.
Plugins
A whole configured setup, bundled and shareable.
Subagents
A fresh window that does the messy work and returns only the answer.
Agent teams
Many windows, coordinated, in parallel.
Hooks
Your code that fires on events and can block, no asking.
Plan mode
Read-only thinking before any doing.
If you can rebuild this list from the three questions (what is in the window, when does it load, what can Claude do), you understand Claude Code.