Deep Dive: Claude Code
Configure, extend, and automate Claude Code until it works the way you already do.
"It worked brilliantly that one time. I just can't make it happen again."
You've been using Claude Code for a while. You wrote a CLAUDE.md. You've read about skills and hooks and MCP servers, and you've probably installed one of each.
But you still watch every tool call go past. You still explain the same project conventions on Monday that you explained on Friday. And the workflow you keep meaning to automate is still living in your head.
That's the gap. Not the model. The setup around it.
What you'll learn
- How Claude Code reasons, and what model and effort settings actually cost you
- How to configure at scale: the
.claudedirectory, modular CLAUDE.md files,.claude/rules/, sessions and checkpoints - How to build custom Agent Skills with arguments, bash, and file inclusion
- How to intercept the execution lifecycle with pre-tool, post-tool, and prompt hooks
- How to orchestrate the built-in subagents and write your own
- How to configure MCP servers, and where the sharp edges are
- How the permission model works, how to sandbox bash, and what prompt injection looks like in practice
- How to hand off work autonomously and verify what came back
- How to run Claude headless in CI/CD, including GitHub Actions
- How to package everything you build as a plugin your team can install
By the end you'll have your own plugin: a skill, a hook safety net, a review subagent and a wired MCP server, packaged into one unit, installed from a clean clone, and published.
81 lessons. Just under seven hours.
Most people stop at CLAUDE.md
It's the obvious first move, so everyone makes it. Then the file grows to four hundred lines, half of it goes stale, and the agent still asks about things you documented in March.
Configuration isn't one file. It's memory, rules, settings scopes, and skills that load only when they're needed. Once you see how those pieces fit, the file gets shorter and the sessions get better.
Autonomy is only useful if you can check it
Here's the trap. You give Claude a bigger task, it works for twenty minutes, it comes back with a diff, and you skim it. It looks fine. You merge.
Skimming isn't verification. It's the same instinct as chasing 100% coverage, arriving from the opposite direction. One measures everything and proves nothing. The other measures nothing and feels fine about it.
So we spend a full module on verification, dynamic workflows, and goal-driven autonomy. You'll learn how to set up work that reports back in a form you can actually trust, and how to drive Claude remotely once you do.
This is not a list of features
The docs already list the features. They're free, and they're more current than any video will ever be.
This course is about the decisions the docs don't make for you. When a hook beats a skill. When a subagent is worth the context it costs. Which MCP servers earn their place, and which ones just add attack surface. Where to draw the permission line so you stop clicking Allow without reading.
By the end, those decisions collapse into one line you keep:
Prompt it once. Skill when it repeats. Hook when it must always happen. Subagent when it needs its own context. MCP when the capability lives outside. Plugin when it should travel.
And a second one, for how you run it. Supervise by default. Automate what verifies itself. Hand off what has a measurable end state. Go remote or headless when you're not there.
You still do the engineering. This is about building the setup that stops you repeating yourself.