Claude Code
What it does
Delegates coding tasks to Claude Code, an AI-powered coding agent that can write, debug, refactor, and explain code across languages and frameworks.
Setup required
None. Claude Code is a bundled skill that works out of the box. Just ask your assistant to write, debug, or refactor code and it automatically delegates to Claude Code in a sandboxed environment.
Permissions
- May require host permissions if working with files on your machine
- Sandbox execution for code testing and evaluation
Common prompts
| You say... | What happens |
|---|---|
| “Write a Python script that renames all files in a folder” | Generates a working script |
| “Debug this error: [paste error message]” | Analyzes the error and suggests fixes |
| “Refactor this function to be more readable” | Rewrites code for clarity and best practices |
| “Explain what this code does” | Plain-English walkthrough of code logic |
| “Write tests for this module” | Generates unit tests |
| “Convert this JavaScript to TypeScript” | Translates between languages |
| “Help me build a REST API with Express” | Scaffolds a project with guidance |
| “Review my pull request” | Analyzes code changes and gives feedback |
What it can do
- Write code in any major language (Python, JavaScript, TypeScript, Go, Rust, etc.)
- Debug by analyzing error messages, stack traces, and code context
- Refactor for readability, performance, or best practices
- Explain complex code in plain language
- Test by generating unit tests, integration tests, and test fixtures
- Scaffold entire projects with file structure, dependencies, and boilerplate
Configuration
Claude Code uses Claude Sonnet by default. It supports scoped worker profiles to control access: general (full access), researcher (read-only), coder (read/write/execute), and reviewer (read-only, analysis-focused). Your assistant picks the right profile for the task automatically
Tips & gotchas
- Give context. “Write a function” is fine. “Write a function for our Express API that validates user input and returns a 422 for bad requests” is much better.
- Paste errors directly. When debugging, paste the full error message and relevant code. The more context, the better the fix.
- Iterate on solutions. “That works but it's too verbose” or “Can you use async/await instead?” Your assistant refines based on feedback, just like a pair programmer.
- File access: If you want your assistant to read or write code files on your machine, it'll need host file permissions. It'll ask before accessing anything.
- Sandbox first: Code is tested in the sandbox before being suggested for your actual project. Safe by default.
That's the full Skills Reference. Ten skills, documented consistently, ready for use. Each one is a starting point. As skills evolve and new ones are added, these pages grow with them. If a skill is missing from this list, it might be in the catalog waiting to be installed, or you can build your own.