The only Claude Code tutorial you’ll ever need (Apr 2026 edition)
A comprehensive tutorial on Claude Code covering setup, core concepts (claude.md, skills, commands, agents, MCP servers), and planning frameworks. The instructor draws on 300+ hours of hands-on experience to guide viewers from installation through advanced multi-agent workflows, using a content repurposing engine as a running example.
Summary
The tutorial opens by framing Claude Code against familiar browser-based AI tools like ChatGPT and Claude.ai, arguing that those tools 'stop at the chat window' — they can think and write but cannot actually execute tasks on your system. Claude Code, by contrast, lives on your computer and can write files, run commands, create folder structures, and interact with apps directly. The instructor shares real community members' results: websites built in a day and a half, landing pages in 15 minutes, and full business apps completed in three weeks.
The setup section walks through both Mac and Windows installations, covering VS Code, Git, Claude Code itself (via a curl command), Node.js/npm, and environment variable configuration. A brief markdown primer explains why .md files are used throughout — LLMs are trained on markdown structure and can parse headings, bullet points, and sections reliably.
The first live demo builds a minimal content repurposing engine in under five minutes using a single prompt: drop in a transcript, get a LinkedIn post, X thread, and newsletter draft saved as separate files. The instructor notes the output was 'more impressive than expected' but lacks voice, structure, and repeatability — setting up the rest of the course.
Permissions are covered next, distinguishing three approaches: the 'nuclear' --dangerously-skip-permissions flag, the default ask-for-everything mode, and the recommended middle path of a settings.json file that pre-approves safe operations (reading/editing files, git, running dev servers) while blocking dangerous ones (package installs, file deletion, sending external data, reading secrets).
The claude.md file is described as 'the most important file in your project' — a session-level system prompt Claude reads before every conversation. A five-question framework structures it: what is this project, how do I run things, what patterns do I follow, what are the gotchas, and what is the workflow. The instructor stresses keeping it under 200 lines and using 'pointer not dumper' logic — reference external files rather than embedding all content inline. Parent inheritance is explained: Claude walks up the folder tree and combines all claude.md files it finds, with closer files taking priority over parent-level rules.
Slash commands are introduced as repeatable, manually-invoked prompts stored as markdown files in a .claude/commands folder. They are good for regular tasks but limited to a single file and require explicit invocation. Skills are presented as the more powerful successor: stored in .claude/skills/[skill-name]/skill.md, they are auto-invoked by Claude based on a YAML front-matter name/description/trigger block. The instructor details 'progressive disclosure' — the description (~100 words) is always loaded, the skill.md body (~200 lines) loads only when triggered, and reference files (platform formatting, voice guides, examples) load only at the specific step that needs them. This architecture prevents context rot.
Hooks (automated actions in settings.json, e.g., ban-word scanners, GitHub staging, notifications), plugins (bundles of skills/hooks/commands installable in one line), and MCP servers (model context protocol — a 'USB-C for software' connecting Claude to Notion, Gmail, Airtable, HubSpot, etc.) are each covered. An MCP demo shows Claude pulling a YouTube video transcript from a Notion database, running it through the content repurposing skill, and writing the LinkedIn post back to the Notion table.
Planning frameworks are matched to project size: built-in plan mode (shift+tab) for tasks under an hour, a PRD generator skill (prd-taskmaster) for multi-hour projects, and the GSD (Get Stuff Done) framework for multi-day application builds. Each creates a written plan file (spec.md or road map) that survives context compaction.
Agents and sub-agents are explained: every terminal window is effectively an agent; Anthropic provides built-in plan, explore, and general-purpose sub-agents; custom agents can be built similarly to skills. Agent teams (still in research preview) allow teammates to communicate with each other via a shared task list. The instructor's controversial opinion is that well-structured claude.md files and skills replicate most agent functionality without the overhead.
The final section covers 2026 features: /loop for recurring scheduled tasks, git worktrees for parallel isolated feature branches, remote control (QR-code phone access to a running session), and conversation resume (claude --resume). The course closes with next steps: write a claude.md, build one skill for a weekly task, connect one MCP, and only then experiment with agents.
Key Insights
- The instructor argues that browser-based AI tools like ChatGPT and Claude.ai all 'stop at the chat window' — they can think and plan but cannot take actions on your system, whereas Claude Code eliminates that wall by living on your computer and executing file writes, folder creation, and app interactions directly.
- The instructor introduces 'context rot' as the core failure mode of Claude Code: as a session grows longer and the context window fills up, Claude makes decisions based on context it can no longer reliably recall, causing quality degradation — and the fix is not asking Claude to 'be better' but structuring tasks so context is fed in small, focused chunks.
- The instructor describes 'progressive disclosure' as the critical architectural principle behind well-built skills: the YAML front-matter description (~100 words) is always loaded into context, the skill.md body (~200 lines) loads only when the skill triggers, and reference files load only at the specific workflow step that needs them — allowing Claude to offload context between steps and avoid bloat.
- The instructor argues against installing skills or plugins globally, stating that globally installed items are hidden in a dot-folder (invisible in the file tree), load into every project whether relevant or not, and create a 'black box' where you no longer understand why certain behaviors are activating — recommending always installing locally first for visibility and debuggability.
- The instructor's controversial position on agents is that a well-structured system of claude.md files with parent inheritance effectively replicates what Anthropic calls 'agents' — each project folder with its own claude.md acts as an agent context, and skills serve as the standard operating procedures those agents invoke, making explicit agent configuration unnecessary for most business users.
Topics
Full transcript available for MurmurCast members
Sign Up to Access