Introduction

This will be a pretty short article. I think most of the people who use AI tools have seen how inconsistent even repetitive work can be. It definitely takes more than one prompt.

Here's how it usually goes with Cursor or Claude Code:

1. You're asking to add auth to the website.
2. Cursor presents you with a list of some libraries, you don't know if they will work or not.
3. Sometimes it will install a package, sometimes add a sign in button, or use a deprecated method, you never know.
4. You're left with half backed integration and having to read the documentation to figure out what's going on and how to finish it.

How to fix it

Add a documentation to context

You can use Context7 or Nia to add up to date information about the library.

Does it help? Partially, it prevents the LLM from using outdated methods, but some prompts they're adding are a mess and model might not use that. We can do better.

Use SourceWizard

I've decided to build an extension of Context7/Nia approach. SourceWizard MCP will search for packages in its library, weighting them by repository compatibility. Once it finds the package it will run the integration in a sub-agent, which will use use a specially written and tested prompt that tells the agent to add everything that is needed for complete integration. For example for auth it will:

• Add the package;
• Add middleware;
• Add sign-up/sign-in pages;
• Add sign in button on the main page;

This works pretty consistently on my tests! Moreover, I can improve the prompts or the library based on previous installation runs.

Click here to try it out yourself.

You will need to install the SourceWizard MCP to your coding agent. And then to install the package just ask something like "Use sourcewizard to add auth", and it will run the integration. You can follow the integration progress via the CLI:

npx sourcewizard@latest status

Let me know how it performs on your project!