Skip to content

What I've Learned from 300+ Hours of Claude Code

Skills, plugins, agentic workflows, and MCP, exploring how a CLI-based AI coding assistant changed the way I build software.

What Claude Code Actually Is

Claude Code is Anthropic's CLI tool for AI-assisted development. Unlike chat-based AI tools, it lives in your terminal with full access to your filesystem, your shell, your development tools. It reads files, writes code, runs commands, searches codebases, and manages git, all inside a single conversation.

I have logged over 300 hours with it. In that time I built an entire social media platform, a real-time stat tracker, this portfolio site, and dozens of smaller tools. All solo. All shipped.

Skills and Plugins

The most underutilized feature of Claude Code is its extension system. Skills are .md files that teach Claude new workflows. Plugins bundle those skills with agents, hooks, and integrations into something you can share and distribute.

A skill is deceptively simple: a YAML frontmatter block followed by markdown instructions.

---
name: deploy
description: Build, deploy to Firebase, commit, and push
disable-model-invocation: true
---

1. Run npm run build
2. Deploy to Firebase Hosting
3. Commit with a descriptive message
4. Push to remote

Type /deploy and the entire workflow runs. But it goes deeper than that. Skills can inject dynamic context through shell commands, accept arguments, restrict which tools Claude can use, and run in isolated subagent contexts.

This is where Claude Code stops being "AI autocomplete" and starts being "AI engineering." You are not just prompting. You are building reusable, composable automation that makes the AI better at your specific workflows every time you use it.

MCP: The Model Context Protocol

MCP is what makes Claude Code genuinely dangerous in the best sense of the word. It is a protocol connecting Claude to external tools and services (Firebase, Figma, Playwright, Google Drive, databases) through standardized server interfaces.

In practice, that means Claude can deploy your site to Firebase Hosting, read your Figma designs and generate matching code, spin up a browser to test your UI, query your Firestore database, and manage your Google Drive. One conversation. No context switching.

The site you are reading right now was built, deployed, and verified through MCP integrations. Claude wrote the code, started the dev server, took screenshots, and pushed to production. I never opened a browser.

Agentic Workflows

Claude Code can spawn subagents: specialized AI instances running specific tasks in parallel. This is where the tool stops feeling like a tool and starts feeling like a team.

  • Explore agents search your codebase to understand patterns and architecture
  • Plan agents design implementation strategies before writing code
  • Code review agents check for bugs, security issues, and style violations

A typical session: I describe a feature. Claude enters plan mode, explores the codebase, designs an approach. I approve. It implements across multiple files, fires up the dev server, verifies with screenshots and accessibility snapshots, and presents the finished result. The whole loop takes minutes.

What I've Built

MooringSocial: A full anti-algorithm social media platform built solo from scratch. Next.js, Firebase, real-time feeds, authentication, media uploads. The kind of project that would normally take a team. I shipped it alone.

Spike Squad: A real-time volleyball stat tracker for my daughter's 2nd grade team, built in a weekend. Claude Code handled the data layer. I handled the UX.

This portfolio site: Built and iterated live in a single Claude Code session. Every section, every deploy, every commit ran through the CLI. Including the blog post you are reading right now.

Patterns That Work

Three hundred hours teaches you things no tutorial covers.

  1. Plan before you build. Use plan mode for anything non-trivial. Getting alignment on the approach before writing code prevents wasted effort.
  2. Build custom skills for repeated workflows. If you're doing it more than twice, make it a skill. Deploy pipelines, code review checklists, project scaffolding. Automate the predictable parts.
  3. Use MCP integrations aggressively. The more context Claude has about your actual environment (databases, hosting, design files), the better its output.
  4. Verify, don't trust. Claude Code can run your dev server, take screenshots, and check accessibility snapshots. Use these verification loops instead of manually checking everything.
  5. Think in systems, not prompts. The real leverage isn't in writing better prompts. It's in building skills, plugins, and workflows that compound over time.

Why This Matters for AI Engineering

The gap between "using AI tools" and "building AI systems" is closing fast. Writing a Claude Code skill is programming an AI agent: defining its capabilities, constraining its behavior, composing workflows that get smarter over time. That is AI engineering, and it is happening right now in markdown files.

The developers who thrive will not be the ones resisting AI tooling, nor the ones blindly accepting its output. They will be the ones who learn to build systems on top of it, treating AI as infrastructure to be configured, extended, and composed. That is the work I do every day.