Works on My Claude

There was a time when the most annoying sentence in software was “works on my machine”. Then containers happened and we more or less solved it. Progress.

We now have a worthy successor: “My Claude knows for sure”.

Here is how a conversation goes in 2026. You ask a colleague a question. The colleague pastes your question into their AI assistant, waits a few seconds and reads you the answer, often verbatim. And that answer is now canon, not a hypothesis or a “here’s what my model thinks” but the actual truth of the matter, which means that if you push back you are no longer disagreeing with a person. You are disagreeing with the Oracle (not the company, though give it time and this one will find a way to charge per answer too). And the person will defend the Oracle with a passion they never once showed for their own opinions.

Never mind that the model may have been missing the three paragraphs of context it needed to get to the right conclusion or that one sentence in there is a perfectly confident hallucination. The part I find funny is different. Even when two people use the exact same model, they almost never use the same harness. Different system prompts, different tools, different memory, different temperature, different everything. Two people running “the same Claude” are roughly where two developers were running “the same Node” on two laptops in 2012. We didn’t fix the problem. We rebuilt it one layer up.

So we get a new kind of friction. Not person vs person, but AI-by-proxy vs AI-by-proxy, with two humans attached who still have feelings about it. And humans being humans, it escalates. Someone takes it personally. Someone agrees in the meeting and complains about it in the DMs afterwards. Neither of them actually holds the opinion they are fighting over. The hurt feelings, though, are entirely their own.

Here’s an idea. If we’ve already outsourced the thinking, why not outsource the arguing too? Let my assistant talk to your assistant. They’d exchange context politely, point out what the other one was missing, and nobody’s feelings would get hurt because nobody with feelings would be in the room. Honestly, it might be more productive. The models are much better at saying “you’re right, I was wrong” than we’ve ever been.

Which leaves one small, awkward question. If the assistants ask the questions, form the opinions, defend the opinions and settle the disagreement, what exactly are we here for?

I don’t have an answer. I asked my Claude, obviously. It had a very reasonable one.