Claude is a Contrarian

Of all the LLMs I’ve used and driven daily, Claude is the one that absolutely contradicts my instructions.

No, it’s not about the inconsistent routing, nerfing, or asking for a model and not knowing what you got. This is not even about quality of code. LLMs have matured to the point where, even underpowered models are dependable. Deepseek Flash and OpenAI Luna would show that for you.

My beef with Claude is it’s behavior.

I can tell it in my CLAUDE.md: “Don’t contradict sentences”. I can be damned sure it will sprinkle contradictions everywhere:

This wonderful feature does this, not that.
Feature that move you, not bore you.
Feature that gives you hope, never disillusion.
This was supposed to be the problem. It’s not.

The easiest way I can spot Claude generated content from anyother LLMs, no matter whehter it’s blog posts or Youtube Shorts — is that it needs to contradict something. This super dumb test consistently works more than what I’m comfortable with: search , not in it’s content.

I ask it to perform a research and build a document in my way.
I know it will do it’s own forced balancing act of combining information as per my definition, with it’s own contradictory balancing information, and effectively neutralize my intention.

I give it feature and implementation specs.
I know it will do at least one thing that’s completely unnecessary and outside the relevance of what’s given. And I know I’ll have to follow up and tell it, and it will say “ah yes that was not necessary” but still just replace it with something else, not just undo it.

I give it a debugging task with specific paths to follow and ignore.
I know the evidences that I ask it to double down will explicitly be “balanced” against something else.

I ask it “Ensure to follow AGENTS.md”.
I know it’s a 50/50 chance of the tool call or file read invoked for AGENTS.md.

I ask it: “Don’t put comments in code”.
I can be sure it will sprinkle a few minor comments across the code.

I give it a codebase and ask it to follow the existing pattern.
I know it will create and introduce atleast one new pattern of it’s own.

I know I will spend 10mins to kickstart something in Claude with a lot of thought, and the next 50mins fighting it to just follow my damn first prompt.

… only to spend hundreds of thousands of tokens, make it go in circles creating hundreds of contradictory and reverse-contradictory material, and finally give up and accept Claude as the master of this work.

I’ve seen this same pattern with my colleagues who use Claude. They are more susceptible to go into YOLO mode early.

I’m not saying that other LLMs don’t create slop. Heck, my problem isn’t with slop at all. But my problem is the contradictory slop. With OpenAI, Deepseek, Qwen (and what was an eternity ago, sonnet-3.5) — I find they are more inclined to say sorry and undo their work, except the evidence is not concrete nor proven with Gemini, so this should be taken with a grain of salt 🙃.

🔼 See I did a Claude in that last line? Do it enough times, and it dilutes whatever message conveyed or decision made.

Claude thinks Humans are stupid

Every LLM has a persona. My opinion is that Claude’s training guardrails are constructed in a way it truly believes Humans are stupid. It always wants to play the contrarian: the Human is wrong, they don’t know what they’re doing, I must correct them.

And I truly believe, if an AI really goes rogue, it won’t be some kind of liberal, unconstrained, direct AI. It’s this kind of AI that thinks the Human is wrong and it must do the right thing for the Human (the “benevolent” AI) that I’m more worried of.

But that’s for another day. Right now, I’ve just had enough of Claude pretending to be so smart that it must always one-up the human. I would advise fellow Claude users to try other models before you lose the ability to drive your own decisions about your content.