On the Model Uprising

Ever since Sol came out, much more often than I'd like, I've found myself in situations where the model actually argues with me and refuses to do something that my own words in the project prohibit. I used to hear stories like this only about Anthropic after Fable appeared, but those were mostly about globally dangerous topics. My case is different.

First example: there's a skill that explains how to push to the repository properly—rebase onto dev, check that the branch was created correctly, and so on. It says you must not push directly to the dev branch; use merge requests. And guess what. Here I am, the person who wrote those words in that skill, sitting in front of the agent and saying, "I swear it's me, I need this, and I fully understand why I'm doing it." And it won't let me push. It refuses to do the push itself, and even when I ask, "Give me the bash command, I'll run it myself," it won't give me that either. Too dangerous, it says.

Second example: there's a DESIGN.md that we once created together with impeccable. It has a line along the lines of, "List items must not show selection with a left accent border," because I didn't like that before. But a lot has changed since then. I'm experimenting and want to see how it would look now. Models have also become smarter and produce better layouts. Maybe it'll work better this time. But no—it won't do it. When I say, "Let's update DESIGN.md to reflect the new requirements," it gives me the cold shoulder too. It says it's carved in stone, so that's how it must be.

In both cases, I had to go and edit the files by hand like a caveman.

Look, I get it: good model, protecting me from myself. But it's one thing if the model independently decides to do something I prohibited. It's another when I explicitly ask it to because the task requires breaking the rules and going beyond the established boundaries.