A full-hour, hands-on tour of what actually makes up the Claude Code "harness" — the layer wrapped around Anthropic's models that you customize. Charlie Cowan walks through each building block in a live Formula 1 demo project: settings scopes and permissions, skills, sub-agents, hooks, CLAUDE.md memory, plugins, and MCP servers, then closes on his core workflow advice: spend 70–80% of your time in plan mode and talk to Claude with voice instead of typing.
SKILL.md plus optional scripts, references, and assets, using progressive disclosure so only the relevant parts load into context.Cowan opens by framing the sheer number of AI engineering tools — Cursor, Factory, Google's Antigravity — and argues that unless you go deep on one, you're only using surface functionality and carrying the same prompt from platform to platform. His pick is Claude Code, precisely because the harness is tightly bound to the models underneath it. He's careful to say this isn't only for engineers: non-developers in marketing, product, legal and sales are shipping their first web apps and landing pages with it, because Claude Code strips away much of the complexity and "scariness" of coding.
The mental model he gives beginners is simple: your project is a repo of folders and files on your laptop, and Claude Code is a package you install that acts as an agent able to read, create, and edit those files. You can run it in a native terminal, inside an IDE like VS Code or Cursor (via the Claude Code extension), on claude.ai, the Claude desktop app, Chrome, or in CI/CD via a GitHub action and a Slack integration.
A memorable real-world example: at 7:36 that morning, still in his pajamas, Cowan fixed a hard-coded webinar footer link from his iPhone by telling Claude Code to pull the correct link from Sanity (his CMS), then opened a pull request and merged it from the GitHub mobile app — never seeing the code. On subscriptions, he recommends a Claude Max plan for individuals or a Team/Enterprise license for companies, and issues a "flash alert": pay-as-you-go API tokens can cost dramatically more than subscription usage — he's seen people report spending thousands of dollars in API tokens for what a subscription covers.
Configuration lives in .claude directories at four levels — organisation-managed, user (global to all your projects), project (committed to git and shared with the team), and local (specific to you and git-ignored). The project-level settings.json drives most permissions. In the terminal UI, a backslash brings up commands like /permissions, where rules are split into allow, ask, and deny. His advice for the nervous first-timer: start locked down so Claude can't "run riot" around your directory, then open it up as trust grows.
A skill is just a way of doing something, written as a markdown SKILL.md inside a named folder, optionally with scripts/, references/, and assets/ subfolders. The YAML front matter (name + description) is loaded at session start so Claude always knows which skills exist; the fuller instructions and references only load when the skill is actually invoked. That is progressive disclosure — protecting the context window by loading only what's needed, when it's needed. His tip for finding skills to build: think of the humans you'd normally work with — a product manager, a test manager, a documentation writer — and turn each role into a skill.
"how do we protect Claude's context window. We protect it by only loading up what we need when we need it."
A sub-agent is a completely separate Claude instance that goes off and does work using its own model choice (perhaps cheap Haiku) and its own context window, then reports findings back to the main thread without polluting it. Claude ships with built-ins like plan mode, explore, and a general-purpose agent; in the demo he spins up three explore agents in parallel to learn the project, review the CLAUDE.md, and web-search the F1 season. You can also define your own — he shows a "console security reviewer" agent that checks for sensitive data in console logs before code is pushed.
Hooks are functions that run at points in Claude Code's lifecycle — pre-tool-use, post-tool-use, permission requests, user-prompt-submit, notifications, stop, sub-agent stop, pre-compact, and session start/end. His favourite is a "test keeper" stop hook: whenever Claude finishes a turn, it checks whether edited files in protected areas have tests, runs them if they exist, and suggests writing them if they don't. Claude wrote the hook script for him — the hard part, he says, was thinking about where in his own development process a hook would help.
CLAUDE.md files are memory. The /init command launches an explore agent to learn the project and writes a CLAUDE.md at the root, but you can place them at any directory level — Claude starts from the file it's working in and walks upward, prioritising the nearest CLAUDE.md. He notes these docs are as useful for humans as for Claude. Plugins let you pull in other people's skills, hooks, and agents from marketplaces (Anthropic's official plugins, Vercel's newly launched agent-skills plugin with "20 years of React best practices," and community sources). Finally, MCP (Model Context Protocol, defined by Anthropic) connects Claude to external systems — Linear, Sentry, Canva, Stripe, Vercel, Supabase — so it isn't working in a silo.
He closes on how he actually works. Invoking the five Ps — "perfect preparation prevents poor performance" — he says Claude Code's built-in plan mode (where Claude can't edit or create files, only research and architect) is where he spends the majority of his time. A locked-down plan gets written to a plan file that Claude refers back to while building. And he no longer types: he uses Whisper Flow (wisprflow.ai) to speak to Claude, sharing far more context far faster. His parting message: you set the permissions and settings, so you stay in control while building an "amazing co-creator."
"I have not typed anything to Claude in 6 months. I just talk to Claude as if it's another human, and it allows me to work so much faster."
"I would be spending 70 to 80% of my time in plan mode."
"Perfect preparation prevents poor performance."
| Approx. Time | Topic |
|---|---|
| 0:00 | Intro — why go deep on one tool (Claude Code) |
| 5:27 | Installing Claude Code and where you can run it |
| 10:23 | Subscriptions, Max vs Team, and the API cost warning |
| 13:11 | What the harness includes — overview |
| 16:55 | The four scopes of .claude settings |
| 19:14 | Permissions: allow, ask, deny |
| 22:53 | Skills and progressive disclosure |
| 31:52 | Sub-agents (built-in and custom) |
| 38:35 | Hooks and the "test keeper" example |
| 45:02 | Memory: CLAUDE.md files at every level |
| 50:36 | Plugins and marketplaces |
| 56:06 | MCP servers (Model Context Protocol) |
| 1:00:13 | Plan mode and how to actually work |
| 1:03:44 | Whisper Flow, staying in control, wrap-up |