Anthropic Claude Code Playbook

Anthropic released a guide this week called How Claude Code works in large codebases: best practices and where to start. It’s aimed at engineering teams running Claude Code on million-line monorepos and decades-old legacy systems — not a beginner audience.

But four of the lessons in that guide apply to anyone using Claude Code, including hobby projects and solo work. Below is what to take from it without needing an enterprise codebase.

What is the article actually about?

Anthropic’s post documents how customers run Claude Code on production codebases that span dozens of repositories, span multiple languages (including C, C++, Java, and PHP), and have decades of history. The takeaway is that the model itself isn’t the variable that makes Claude Code work well at that scale — the setup around it is.

That setup is what Anthropic calls the “harness.” And the harness is where every Claude Code user has direct control, regardless of project size.

Why should beginners care about an enterprise guide?

Three reasons:

• The same setup works at any scale. A CLAUDE.md file in your hobby repo does the same job it does in a 5,000-engineer monorepo — it gives Claude the context it needs to make good decisions about your code.
• The credibility question is settled. If Claude Code works on real, ugly production systems with decades of legacy patterns, it’ll handle whatever you throw at it. The article is the closest thing Anthropic has published to “here is the evidence this is real.”
• The four habits are easy to copy. None of them require a team. They take about an hour to set up and pay off every session afterward.

What is “the harness” and why does it matter more than the model?

Anthropic’s exact wording: “the ecosystem built around the model — the harness — determines how Claude Code performs more than the model alone.”

In plain English: the model is fixed. What you can change is everything around the model — what context it gets, what tools it can call, what conventions it follows, what it’s allowed to do. That collection of things is the harness. The article lists five extension points:

• CLAUDE.md files — instructions loaded automatically every session
• Hooks — scripts that run automatically at certain moments (more on hooks here)
• Skills — specialized expertise that loads on demand (more on the Anthropic Skills Library)
• Plugins — bundled configurations you can share with others
• MCP servers — connections to outside tools (see the Claude API for the underlying protocol)

The point: Claude Code is a system you configure, not a magic box. Every layer of that system is yours to shape.

What goes into a CLAUDE.md file?

A CLAUDE.md file is a plain-text instructions file you put at the root of your project (or in any subdirectory). When Claude Code starts a session, it reads the file and treats it as project context. Anthropic’s guide says to keep these files lean and layered.

Lean means: short. Three to five short paragraphs. Don’t document everything — document the things Claude needs to know that it can’t figure out from reading the code.

Layered means: put a small CLAUDE.md at the repo root with project-wide stuff, and put more specific CLAUDE.md files inside subdirectories. The frontend folder gets a CLAUDE.md about frontend conventions. The infrastructure folder gets its own. Claude picks up the relevant one based on what you’re working on.

Good things to include:

• What this project does (one sentence)
• The package manager, build tool, and test runner (one line each)
• Conventions that aren’t obvious from the code (e.g. “always use async/await, not promises”)
• Past mistakes that should not be repeated (e.g. “don’t edit migrations after they’ve been deployed”)
• Anything you keep having to remind Claude about

Bad things to include: anything Claude can read from the code itself. Don’t duplicate documentation that already exists.

How do skills, plugins, and MCP servers fit in?

Three different extension mechanisms, three different uses:

• Skills are bundles of expertise Claude loads when you need them. Example: a “PDF processing” skill that knows the relevant libraries, file formats, and edge cases. Anthropic publishes a free Skills Library with several dozen of these.
• Plugins are distributable bundles. If you build a useful Claude Code setup for your team or for a domain (WordPress, React, data engineering), you can package it as a plugin and share it. See our guide to Claude Code hooks and plugins.
• MCP servers are how Claude connects to outside tools. Want Claude to read from your database, query your ticketing system, or pull from Notion? An MCP server makes that possible. MCP is the underlying protocol.

For solo projects, you usually don’t need plugins. You can get a lot of mileage from a good CLAUDE.md plus one or two skills.

When should you refresh your Claude Code setup?

Anthropic says to review your CLAUDE.md and skill configurations every three to six months. The reason: instructions that helped an older model can hurt a newer one.

Their example is specific: a CLAUDE.md rule telling Claude to break every refactor into single-file changes might have helped an earlier model stay on track. But a newer model can plan and execute coordinated multi-file refactors safely — and that old rule now prevents it from doing the smarter thing.

Practical advice: every time a new Claude model ships (Sonnet 4.6, Opus 4.7, future versions), spend twenty minutes reading your CLAUDE.md and asking yourself which rules are still earning their keep. Delete the ones that aren’t.

What about the parts that are obviously enterprise-only?

Several pieces of the article are clearly for organizations:

• Assigning a DRI (directly responsible individual) for Claude Code rollout — you don’t need this when you’re the only person on the project.
• Language Server Protocol (LSP) integrations — useful in big monorepos with multiple languages, less critical in a single-language solo project. Worth knowing about, not urgent to set up.
• Subagent architecture for parallel exploration — powerful but usually overkill for a hobby project. If you want to understand the idea, see our subagents glossary entry.

The article’s value for solo users is the framing. The lesson that “your setup matters more than the model” reframes how to think about getting better at Claude Code — it’s not about prompt-engineering harder, it’s about building a better configuration layer.

Frequently asked questions

Do I need to use Claude Code for big projects to get value from this article?
No. The harness concept applies to projects of any size. A solo hobby project benefits from a good CLAUDE.md the same way an enterprise codebase does.

Where does CLAUDE.md actually live?
At the root of your project, named exactly CLAUDE.md (capital letters, with the dot-md extension). You can also put one in any subdirectory and it’ll apply when Claude is working in that folder.

What’s the difference between a skill and a plugin?
A skill is one piece of expertise (e.g. “how to work with PDFs”). A plugin is a bundle of things — possibly multiple skills, hooks, MCP server connections — packaged together so someone else can install your whole setup in one step.

Is Claude Code free?
The tool itself is free to install. Using it requires a Claude subscription or API access. See our Claude review for current pricing.

How does Claude Code compare to GitHub Copilot or Cursor?
They’re different categories. Copilot and Cursor are editor-integrated autocomplete/chat tools. Claude Code is an agentic terminal tool that can plan multi-step changes, run commands, and edit multiple files in one task. See our deeper comparison at Claude Code vs Cursor vs Copilot.

I’m brand new to coding. Should I be using Claude Code at all?
You can, with realistic expectations. It’s very useful for explaining what existing code does, helping you debug, and walking through small projects. The bigger your project gets, the more the harness matters — so start small, build the CLAUDE.md habit early.

Sources

• Anthropic — How Claude Code works in large codebases: best practices and where to start (May 14, 2026)
• Claude Code official documentation
• Anthropic news — current Claude model family

You might also like

• What Is Claude Code? — the beginner introduction.
• Claude AI Review — the model behind Claude Code.
• Claude Code Hooks and Plugins Guide — extending Claude Code with hooks and plugins.
• Anthropic Skills Library — the official free skill catalog.
• Claude Code vs Cursor vs Copilot — how the three compare.
• Best Claude Code Extensions — recommended add-ons.
• What Are Sub-agents? — the architecture pattern from the article.
• AI Glossary — every term in plain English.