@andrewh , here’s how I understand it (and please correct me if I’m wrong — I’m genuinely trying to learn too):
Custom modes weren’t just nicer prompts. They actually changed the environment the agent was working in. Rules don’t do that. That’s basically the root of why people are upset.
From what I’ve seen, you can try to solve things with multiple rules, but rules don’t remove tools — they just ask the model not to use them. And LLMs don’t always listen. If bash or some MCP is loaded, the model eventually tries to use it no matter how many “don’t do this” rules you stack on top. You can add 10 rules and the model will still slip when it gets confused or sees a chance to call a tool.
Custom modes actually disabled specific tools and MCPs. The model literally couldn’t use them even if it wanted to. That’s a completely different level of control. People used that to make read-only research modes, “no bash” modes, or MCP-specific modes where the context didn’t get polluted by a huge list of tools.
Without modes, all MCPs and tools are always visible, and for folks using multiple MCPs, that quickly becomes noisy. The model starts calling stuff at random simply because it sees the tool, not because the user wanted it.
Also, a lot of people were using modes to enforce architecture rules, naming conventions, repo-specific patterns, etc., automatically. Not by writing huge rule lists, but by having a mode that acted like a “specialized worker.” Trying to cram all of that into one global prompt with a bunch of rules just isn’t the same — the model ignores parts of it, merges other parts, forgets half of it after a few interactions.
And switching between tasks was easier too. One mode for refactors, one for documentation, one for backend work, one for read-only research, all with their own toolset and even their own preferred model. Rules and commands can’t swap models or tool availability, so you lose a lot of the control that made modes reliable.
So yeah, when someone asks “why not just use multiple rules?” — I think the answer is basically: rules don’t actually change what the agent can do, only what you tell it to do. Custom modes changed the actual sandbox.
Again, this is just my understanding based on how me and other people were using the feature.
I heard Google has a new IDE.