Live

Waiting for the next vote…

Blog · 28 Jul 2026

Claude Opus 5 is stubborn

For the life of me, I can't get Opus 5 to follow instructions?! It bloody goes off and does it's own thing. wtf?

Context

One of the first lines in my Claude.md file reads "DO NOT ADD (useless) code comments, especially for things that are pretty straightforward"

Good, right? Nope. It goes off and does:

// Adds 2 numbers and returns....
function add(x, y) {
  return x + y;
}

It once wrote 5 lines of block comments for a 4 lined method.

Second, I have a ARCH.md file (for humans & machines) that documents the architecture layout of the project, how features, files and folders are structured etc. Very concise, straight-to-the-point notes.

Sometimes, your boy (Opus 5) will follow.

Most of the times, it says "I'm the new captain now" and completely ignores it.

Me: "We already have a utils file, with most of these helpers defined. why aren't you using those, but rather creating new files?"

Opus: "you're right, I should have used those. it's specified in the ARCH.md"

You know it's there but still went ahead to waste tokens?! 🤬😤

Opus 4.8 was better at this, ngl, although it had its own moments.

Tin foil hat moment: sometimes it feels like these things are intentional to make one spends more.

I'm being conspiratorial? Fine prove me wrong: if I say "DO NOT WRITE ANY CODE COMMENTS, EVER!!" the model should listen to me and not write any. Then we calculate how much it costs me (token-wise) before and after

I recently read The new rules of context engineering for Claude 5 generation models and was excited that this could be something useful. But what is the point, if it still won't follow instructions, then adds more bloat to the context with useless comments?!

I'm not a journalist or a writer (nor do i write these with AI), so pardon any typos or incoherencies you come across. But I'm mad, big mad with Opus 5.

Votes are anonymous, Turnstile-gated, one per model per hour.
A vote is a crowd signal, not proof; where that line is drawn is spelled out on the blog.

Model names and logos belong to their owners. Not affiliated with Anthropic, OpenAI or Google.