This folder contains the necessary setup to get started with Context Engineering in Claude Code. Please follow the instructions to apply the files to your own project.

First, clone this repository to disk:

git clone https://github.com/wmeints/context-engineering/

Then, copy the following files to your project:

claude-code/.claude/commands/generate-plan.md   # Prompt template to create requirements documents.
claude-code/.claude/commands/implement-plan.md  # Prompt template to implement requirements.
claude-code/implementation-plans/.gitkeep       # Folder to store the requirements documents.
claude-code/src/.gitkeep                        # Source folder (optional)
claude-code/CLAUDE.md                           # General instructions for the project

If you're using this on a completely new repository I recommend setting up the project structure using some kind of project template first. For example, you can use create-react-app, dotnet new, quarkus create app, or uv init to start new projects in various languages. Setting up a completely new project structure with an LLM is bad idea.

To start working on the project, make sure to fill in the details in the CLAUDE.md file. You'll need to provide the following:

• Goals for the solution: What is the solution supposed to do?
• Technology stack: What technology stack do you want to use?
• Project structure: How is the project structured?
• Testing the project: How do you want things to be tested?
• Documenting the code: Where and how should details be documented?

I recommend writing short and concise instructions for the LLM. This helps keep enough space available in the context window for actual work.

Tip: Unsure of what to write in the CLAUDE.md file for your project? You can ask Claude Code to generate the file for you. Use the /init command inside Claude Code to get an initial set of instructions based on what your project looks like.

Before you start using Claude Code to write code, you'll need to come up with an initial set of instructions for the requirements generation. Create a new file TASK.md in the root of the repository with the following content:

# Feature

<Describe the feature that you want to implement here.>

## Examples

<Insert examples here. You can provide them inline, or list files located in
the `examples` folder if you have longer examples.>

## Documentation

<Refer to any relevant documentation such as input from a client or online
documentation for libraries you want to use in the implementation later.>

## Other considerations

<Provide other considerations that are relevant to the feature. You can list
things here that the agent is missing a lot.>

I prefer to keep the TASK.md file checked in as a reminder to myself of what I need to provide as input for generating requirements.

After you've created the requirements document, you can start to generate requirements. Start Claude Code with claude and then issue the following command:

/generate-plan TASK.md

It will take a few minutes for Claude Code to write the full requirements document. You can continue to iterate on the requirements manually or by opening Claude Code in the terminal inside VSCode. The agent will automatically focus on the requirements document when you have it open in VSCode alongside Claude Code in the built-in terminal.

The generated requirements document will be stored in the docs/implementation-plans directory.

Once you're happy with the requirements, you can start to generate code.

Use the following command in the Claude Code terminal to kick off the implementation of the generated requirements document.

/implement-plan docs/implementation-plans/<your-plan-doc.md>

This will again take a few minutes to complete. This is a great time to get hydrated, take a bio break, and talk to a colleague.

Once you have the code it's important to verify that the code is to your liking. You can do this by configuring Claude Code as a PR reviewer.

I recommend manually reviewing your code because the AI can't catch every problem in your code.

I recommend taking a close look at the generated requirements and breaking them up into smaller chunks. While Claude Code can write impressive amounts of correct code, it's quite hard for you to review since you'll be overwhelmed.

I typically copy/paste chunks into new markdown files, but of course you can ask Claude Code to split the generated requirements into smaller chunks in a more semantic way.

The more details you can provide about your solution in documentation the easier it will be for Claude to find out how to approach the problem. I recommend writing architecture documentation using Arc42 and storing it in the docs/architecture directory in the project.

Requirements are already stored in docs/requirements to follow this convention.

One of the most powerful features of using an agent is the availability of the Model Context Protocol (MCP). You can connect the agent to a lot more context information through this protocol. Make sure to check out the modelcontextprotocol/servers repository for a collection of MCP servers that can provide additional context information.

Add a CLAUDE.local.md to the project root with personal communication preferences. Make sure you include only personal preferences. Place project-level preferences in the CLAUDE.md file.

This template uses a single CLAUDE.md file in the root but you can add more detail by putting a CLAUDE.md file within your project directories. For example, you can provide specific instructions for your web frontend by placing a CLAUDE.md in the web frontend directory.