Composer 2 almost never does as asked

I’ve gotten to the point where I cannot trust Composer 2 to do anything of value. These are the common behaviours I’ve seen:

  • Ignoring explicit instructions (even when labelled explicit or invariant)
  • Latching on to an old part of the context and applying it to everything that follows
  • Reasoning itself out of the thing you wanted and doing something else entirely
  • Applying a correction to all of its prior actions except the most recent, which is the one you meant.
  • Saying it will apply something in a specific way and then doing something else entirely.

If you want examples, I would suggest a random sample of my interactions with it would yield a 90% hit rate. In my mind, it’s unusable for me.

1 Like

don’t use bad models if you care about good responses?

This. I have noticed specifically over the last few days it has all but abandoned following my Cursor rules. ALL other models follow them to a T.

I have disabled Composer 2. Let’s see if that has any impact…

same