▸ TLDR
CHECK 04 10 MIN · DIAGRAMMED

How Claude Code works

What you’ll know by the end of this check

  • The five-step agentic loop that powers every Claude Code session
  • Why the context window is the single most important budget you manage
  • The three permission modes — and when each one is the right tool

The shortest possible answer

Claude Code runs a loop:

  1. You submit a prompt.
  2. Claude gathers context by calling tools (reading files, searching, running commands).
  3. It takes an action (edit a file, run a test, hit an API).
  4. It verifies the result.
  5. If the goal is met, it stops. If not, it loops.

You can interrupt, steer, or add context at any step.

The three parts that actually matter

Context window

Claude has a finite working memory. Everything it reads, every command output, every message you type — all of it takes up space. When you approach the limit, Claude Code compacts the conversation automatically, summarizing older content to free room. Compaction can lose details, so it’s not free.

The reason this matters: most people don’t manage context deliberately, hit the limit, and watch quality silently degrade. Check 07 is the full playbook.

Tools

Tools are how Claude moves beyond “text in, text out.” File reading, web search, shell commands, editor operations, MCP servers, skills — all tools. Claude decides when to call each one based on your prompt and what it’s already seen.

Permissions

ModeBehaviorUse when
Default (ask)Asks before every file edit and every shell commandUnknown codebases, destructive work, learning the product
Auto-acceptWrites files without asking. Shell commands still ask.You’re iterating fast in a safe branch
Plan ModeRead-only. Builds a plan, doesn’t executeAnything you’d usually think about before doing

Be careful skipping permissions. Giving Claude free rein means a mistake is harder to catch before it happens.

Things to try right now (5 minutes)

  1. Watch the 3-minute diagrammed video at the canonical source.
  2. In Claude Code, run /context to see your current context usage. Memorize the shape.
  3. Cycle through all three permission modes with Shift+Tab. Read the status bar each time.

The canonical version

Full official lesson is at anthropic.skilljar.com/claude-code-101/469789.

Ready to verify this check?

You can draw the agentic loop on a napkin. You know what /context tells you. You know which permission mode you’ll default to. Mark it cleared.