Claude Code Documentation (June 2026)

The Core Framework

• Claude Code is an agentic terminal tool — Claude reads files, runs commands, and works autonomously while you watch, redirect, or step away.
• Context is the scarce resource. Almost every best practice (/clear, /compact, /rewind, subagents, lean CLAUDE.md) reduces to keeping the window clean.
• Scale outward along a ladder: single session → subagents → background sessions / agent view → agent teams → dynamic workflows ("ultracode").
• Safety is layered (defense in depth): permissions → sandboxing → auto-mode classifier → hooks. Only permissions.deny is an unconditional wall; managed settings sit above all of it.
• Verify the work: give Claude a way to check itself (tests, /goal, a Stop hook, /code-review).

Quick Lookup

Situation	Do This	Avoid This
Context filling up	/clear, /compact, or a fresh session	Letting one session sprawl
Big task, one context	Subagents → agent teams → dynamic workflows	Cramming it all into the main thread
Risky/destructive actions	Permission rules + sandboxing + auto mode	Relying on the model to "be careful"
Must-always-happen action	A hook (deterministic)	CLAUDE.md instruction (advisory)
Investigate without polluting context	A subagent (isolated context)	Reading everything inline
Org policy	Managed settings (precedence; arrays merge)	Per-machine local config drift
CI auth	claude setup-token (long-lived OAuth)	Baking in a raw API key
Parallel edits colliding	Git worktree isolation (.claude/worktrees/)	Multiple sessions on one tree

The Key Insight

The context window holds the whole conversation and performance degrades as it fills — so most of Claude Code's features and best practices exist to manage that finite window while scaling work outward. (Core thesis, Best practices section)

Key Diagrams: The scaling ladder · Layered safety model