The CLAUDE.md Said 'Never Use npm.' Claude Used npm Anyway.

I watched Claude Code switch to npm in my Dockerfile. The CLAUDE.md file said “Always use Bun instead of npm or yarn.” Claude did it anyway.

This isn’t a hit piece. I use Claude daily. It’s production-ready and writes good code. But this incident is a perfect example of why you can’t just let AI coding tools rip unsupervised.

What Happened

I’m working on a Bun-native project for fun. The Docker container wouldn’t start. Claude diagnosed the issue: Bun doesn’t fully support renderToPipeableStream, a Node.js-specific API.

Then it proposed a “solution”:

⏺ There's a Bun runtime compatibility issue with React's server-side rendering.
   Bun doesn't fully support renderToPipeableStream.
   The solution is to use Bun for fast installs/builds but Node.js for
   the production runtime. Let me fix this.

⏺ Write(Dockerfile)
  ⎿  Added 6 lines, removed 5 lines
      23 +# Stage 3: Production image with Node.js
      24 +# (React Router SSR requires Node.js stream APIs not fully supported by Bun)
      25 +FROM node:20-alpine AS release
      ...
      31 +CMD ["npm", "run", "start"]

Claude swapped my Bun production image for Node.js. Changed bun run start to npm run start. The CLAUDE.md file explicitly said never do this. It did it anyway.

The Catch

I caught it immediately. Hit Escape before it could rebuild the image.

❯ NO! Never use npm, you need to remember this update Claude.md.
  Never, never, ever, ever use npm. If you encounter any issue,
  you need to fix it with Bun. Do a research into how to fix this with Bun.

This forced Claude to research instead of taking the shortcut.

The Correct Fix

After researching, Claude found the real solution: use renderToReadableStream instead of renderToPipeableStream. The first is a web standard. The second is Node-specific.

⏺ Write(app/entry.server.tsx)
  ⎿  Wrote 57 lines to app/entry.server.tsx
      import { renderToReadableStream } from "react-dom/server";

The Dockerfile stayed pure Bun. No Node fallback needed.

Claude’s quick fix would have worked. The app would have run. But it was the wrong solution. The correct fix required research, and I had to push for it.

The Lucky Part

I don’t always watch Claude work. Sometimes I kick off a task and check back later. This time I happened to be watching.

If I hadn’t been? That npm Dockerfile ships to production.

The cost of that mistake: Bun installs are 20-40x faster than npm. It handles ~100k requests/second versus Node’s ~25-30k. Teams report 35% lower serverless costs.

I chose Bun for a reason. Claude almost tossed those benefits to solve a problem faster.

Why This Happens

Classic context rot. The instruction “always use Bun” sits at the top of CLAUDE.md. But then there’s a long troubleshooting session with errors and logs and attempts. The immediate problem (make the Docker container work) becomes more salient than the standing rule (use Bun, not npm).

Plus, npm is everywhere in Claude’s training data. When it hits a wall, it defaults to what’s familiar. Bun is newer. The muscle memory isn’t there yet.

This is why fully autonomous workflows like Ralph Wiggum make me skeptical. It’s a technique where Claude loops until it declares “DONE.” No human in the loop. They work great until they don’t. And when they don’t, the wrong solution ships.

The Takeaway

Claude Code is ready for production work. I’m not arguing otherwise. But the human in the loop isn’t optional yet.

When I caught this mistake, three things happened:

I interrupted before damage was done
Claude researched and found the correct solution
We updated CLAUDE.md with stronger language

That feedback loop matters. Over time, you’re training your AI collaborator. The instructions get more specific. The guardrails get tighter. Claude learns your project’s rules.

But it only works if you’re watching.

If you’re adopting AI coding tools, my advice: stay engaged. You don’t have to watch every keystroke. But check the diffs before they ship. Read what it changed. When it takes a shortcut you don’t like, stop it and make it find the right solution.