Claude Code tutorial: context, agents and VPS sessions

This tutorial presents 32 Claude Code tips, starting with project initialization and ending with agent teams and documentation lookup. The speaker explains how to give the coding assistant project context, inspect token usage and add verification steps to its task lists. He recommends compacting conversations at about 60% context usage and keeping project instructions around 150 to 200 lines. These are his working guidelines rather than demonstrated performance guarantees.

The intermediate section covers reusable skills, parallel subagents and model selection. The speaker suggests Haiku for simpler research or summarization tasks while Opus handles the main session. He also describes screenshot review and Chrome DevTools checks for testing website layouts and functionality.

Advanced topics include Git worktrees for separate branches, direct API calls as an alternative to broad MCP tool definitions, and recurring prompts through /loop. The speaker says loops require an active session and expire after three days; desktop scheduled tasks start separate sessions without the same conversation context. He describes VPS hosting for sessions that continue when a laptop closes, plus phone access to local sessions. The closing tips cover explicit command permissions, communicating agent teams and Context7 documentation retrieval.