AI Ready: Advanced Claude Code Lab

Table of Contents

1. Challenge

   Introduction

   Claude Code offers a number of extensions and advanced features that let you more effectively use Claude with your coding projects. These include the ability to connect to Model Context Protocol (MCP) servers, use skills to consistently perform repetitive actions, work with subagents to perform specialized tasks, leverage git worktrees to perform tasks in parallel, and use hook events to trigger actions before or after certain tools are used.

   A Python-based MCP ticketing server is provided and already running in the lab environment. This server contains a number of tickets that track bugs and other issues you (that is, Claude) need to resolve throughout the duration of this lab.

2. Challenge

   Claude Code and Model Context Protocol Servers

   MCP servers work as an adapter between your AI agent, such as Claude, and an external data source – in this case, a ticket tracking service that records bugs and tasks that need to be resolved within an application.

   Without an MCP server, to work with an existing data source, Claude often needs to leverage a wide variety of tools, tool schemas, and functions to perform the required task. An MCP streamlines this process, working as an integration layer that provides Claude with everything it needs to work with the external service. MCP servers also provide the benefit of standardizing how Claude will work – Claude will not need additional context to rediscover how to work with the server every time it needs to work with the external data source; it just uses the tools provided in the connection layer to perform the task(s).

   To connect to an MCP server use the claude mcp add command, providing a server name and details about how to connect to the server. For this lab, a local ticket server is provided via Python script. To launch it, you'll reference Python via the virtual environment with "$(pwd)/.venv/bin/python" and the server itself with "$(pwd)/server/ticket_server.py". Since Claude is what is launching the server, using absolute pathing ensures everything will start correctly. By default, Claude Code does not show results from the MCP tool that are longer than 50,000 characters. Instead, it saves the text to a file and shows a preview of the results. You can alter this behavior through the use of the MAX_RESULTS_CHARS setting. To change this or any setting for an MCP server, use the -e flag when connecting the server. Up until this point, you have configured the MCP server to work with a private, local entry – notice the --local in the previous claude mcp add commands. To have the MCP server work for anyone who works with the project, you can set up the scope so it's project-level. This creates a .mcp.json file in the project directory that will be available for anyone who clones the project.

3. Challenge

   Standardizing with project instructions and skills

   Most often, you will not be the sole person working on a project, and you will not be the only person using Claude with that project. To ensure Claude works consistently across shared projects, you should do things like provide project instructions in the CLAUDE.md file, which is read by Claude whenever started within the project; and provide reusable Claude skills that consistently perform common tasks within the project itself.

   The project in this lab already has a CLAUDE.md file with basic information about the project. The configurations for skills are also provided in the ticket-server/claude/skills directory; these are not in the appropriate spot for Claude to use them and are not yet enabled. To add reusable Claude skills to the project, create and define a skill in a SKILL.md file within a .claude/skills/<skill-name>/ directory. Skills can be set up so Claude only uses specific, defined tools and commands; and skills can be configured so they can only be run manually by the user or Claude can automatically run it when it thinks the skill will be appropriate. The fix-ticket skill requires the user to manually run it for Claude to use it. In contrast, you can also configure skills so Claude will use them when it deems appropriate without the need for a user to manually run the skill. This was disabled via the disable-model-invocation parameter in the previous skill, but is otherwise enabled by default.

4. Challenge

   Delegating work with subagents

   Subagents are specialized AI workers that can be run in parallel with your regular Claude session. These agents are provided with instructions and limited tools to perform their tasks, and have their own context and context window. Once a subagent completes its task, it sends a report back to Claude.

   The primary difference between skills and subagents is this use of context: skills work within the context of the current conversation; whereas subagents only have the context that is provided within the subagent configuration, and therefore always work with fresh context. This isolates the work and can limit token spend.

   Subagents are configured for a project in the .claude/agents directory, and, like skills, are configured within a markdown file.

5. Challenge

   Using git worktrees

   A git worktree lets you check out multiple git branches in a single repository in parallel by working in different directories. When paired with Claude, it means Claude can run two or more sessions at the same time, and resolve multiple issues at once without touching the files of another session. To use worktrees with Claude Code, use the --worktree flag to define the name of the desired worktree. The worktree itself will be created from the latest commit.

6. Challenge

   Automating tasks with hooks

   A hook is a script or command that Claude runs automatically before or after certain actions are performed. Hooks are ideal for enforcing project rules, automating tasks, and integrating Claude Code with existing workflows and standards.

   A PreToolUse hook is set to run before a tool is used, and can even block the use of certain tools. PostToolUse hooks run after a tool is used. Hooks are defined in the Claude settings using JSON, and given that they contain code and commands that affect how Claude works with your greater project, should always be tested before being added. Hooks also expect to receive a JSON description of the action in standard input to perform its action. The hooks are now set up for use locally—notice the .local in the Claude settings file. Hooks, like skills, can also be shared with the greater project using the project's shared settings at .claude/settings.json. Before sharing, however, you should test the hooks with Claude itself and not just the setup script.