The core value of Custom Modes was controlling the LLM experience without relying on the LLM. For one example, manually disabling an MCP is not the same as telling an LLM to “please not use it” in a Command’s Markdown file.
Additionally, a mode and all its customizations could be set once and persist across repeated prompts and Agent conversations. A Command must be manually typed again every single time. This makes it much more likely for the human to forget to type it and waste time (and tokens) on an incorrect prompt.
A key differentiator of Cursor versus other AI-assisted coding tools has long been the rich deterministic tuning available that takes place before LLMs start processing tokens. Examples that come to mind include Rule activation criteria, the context pills… and Custom Modes.
Please bring them back; I cannot achieve with Commands what I was doing with Custom Modes.