Custom commands are not an adequate replacement for Custom Modes - ideas to solve

Ideas Feedback
1

Love Cursor, but I really depended on Custom modes to have explicit control over:

  • The specific model used for the task at hand
  • Whitelisting/blacklisting tools that it could use
  • Specifying system instructions and file access

I cannot specify a specific model for the command. The permissions are not respected after a few turns/basic context rot - making this wholly unusable for development tasks in a highly regulated (healthcare) environment.

We depended on several ‘read only’ custom modes to assist us during development etc.

Examples are:

  • Browser debugging “Investigate bug in UI, but don’t actually change any code after you find the problem”:. At least with custom modes it would be blocked if it did attempt changes - now my command just blows right through and trys to fix what it finds….

Maybe I am not using it right, but i dont have time to prompt engineer where I could previously just pick a checkbox

At this juncture, I just use Claude & Kiro/Q with the --agent command, for when I need something done with a custom agent.

—- How would this be solved? —-

Just add back params for model and allowed_tools and a few other controls maybe? Doesn’t seem like a very hard problem to solve but again - maybe I am just holding things wrong.

2

Hey, thanks for the detailed feedback. I fully understand your situation with compliance requirements in healthcare.

You’re right: Custom Commands currently aren’t a full replacement for Custom Modes when it comes to explicit control over tools and models. The team is aware of this limitation and is working on improving the experience.

The current official guidance is to use text instructions in Custom Commands (Modes | Cursor Docs), but as you pointed out, that doesn’t guarantee the restrictions will be followed.

It would really help the team if you could share a few concrete examples:

  • Chat request IDs where the agent ignores the restrictions after a few turns
  • The exact instructions in your custom commands that aren’t being followed
  • Screenshots showing where the agent breaks through the restrictions

That will help us prioritize work on an explicit permissions system for Commands.