CLAUDE CODE

Explore and propose changes without executing them.

Eliminate prompts with a background classifier that judges safety.

Only pre-approved tools allowed. Fully non-interactive for CI.

Skip all permission checks. Container and VM use only.

Granular allow/ask/deny rules per tool with wildcard patterns.

Auto-guarded directories like .git, .claude, and .vscode.

Filesystem and network isolation for Bash tool calls on Linux and macOS.

Use opus, sonnet, haiku, and best to switch models easily.

Hybrid mode: Opus for planning, Sonnet for execution.

Extended context window for Opus and Sonnet on supported plans.

Low, medium, high, xhigh, and max for adaptive reasoning control.

Interactive UI to switch models and effort sliders mid-session.

2.5x faster Opus at a higher token cost (research preview).

Toggle with Alt+T. Claude reasons through complex problems before responding.

Add gateway or custom models to the picker via environment variables.

Reusable markdown files with instructions and workflows.

/simplify, /batch, /debug, /fast, and other built-in skills.

Trigger with /skillname or let Claude auto-load when relevant.

Configure model, effort, tools, MCP servers, and invocation scope.

Pass arguments to skills with string substitution support.