[vunderba]

Related but I decided to try out Claude Code and one of the first things I noticed is that it didn't seem to be checking for basic transpilation errors (this was in a Typescript project). I added the following bit to the CLAUDE.md file (which is the equivalent of the assistant's LLM guidelines):

   - `npm run build` - Run TypeScript compilation + Vite build. You should periodically run this to check for errors.
 

And it made a pretty significant difference when "autotake suggestions" was on.

[anon7000]

Cursor does have a loop with Claude which fixes lint & type issues which is nice

[arthurgousset]

Oh interesting, thanks for sharing! First time I hear about the `CLAUDE.md` file. Catching transpilation errors seems like a great use of custom instructions.

Do you have any other custom instructions in there?

[vunderba]

So my basic approach is:

1. Create the project folder

2. Initialize everything using a template scaffolding system (vite, nextjs, etc.)

3. Create a simple README.md which contains a list of the desired tech stack (for reinforcement purposes) and goal of the project.

4. Start up Claude

5. Use the `/init` command which creates the vanilla CLAUDE.md file

Interestingly, Claude captures all the NPM commands from the package json into the instructional file but the issue with this is that in my particular test case it kept trying to run `npm run dev` which spins up the app itself. That's... not ideal since I'm running it myself so it can hot-reload and I can see the changes applied immediately. I usually delete that line.

But I should note that I've never used any "agentic coding assistants" before (Aider.chat, Cursor, etc.) - so I'm sort of learning as I go.

One thing I'd be curious about is if there is a "clientside browser" MCP that Claude (or other assistants) can monitor - because ideally you'd really want Claude to be doing the following periodically:

- Taking a screenshot of the browser and using some multimodal system to sanity check

- Monitor the Dev Tools console for errors/warnings/etc.

- Read through the DOM to sanity check

So I have "npm run dev" already running and I could theoretically instruct Claude (or another system) to attach itself to the Chromium/Firefox browser window.

This MCP server seems to be focused on Node/serverside debugging.

[arthurgousset]

Thanks for the detailed write-up, this is super useful! I really appreciate it.

> Claude captures all the NPM commands from the package json into the instructional file

Oh neat, I didn't know that. I'll try to use `/init` next time!

> it kept trying to run `npm run dev` which spins up the app itself. That's... not ideal since I'm running it myself so it can hot-reload

+1 I regularly notice that behaviour too. I'm running node in dev mode, then Cursor wants to run node, but the port is already in use.

> I'd be curious about is if there is a "clientside browser" MCP

Great point, I noticed this is a common feature request for front-end development. I haven't used it myself, but multiple people mentioned BrowserTools [1]. I think it can read the Dev Tools console, but I'm not sure about multimodal or access to the DOM.

[1]: https://github.com/AgentDeskAI/browser-tools-mcp