I’ve been playing with both Claude Code and Cursor agent a ton. One feature I love with Claude Code is the /compact command when the context starts to get unwieldy. With cursor agent I tend to typically just start new sessions when things get unwieldy, but sometimes i really want to keep the context of what i’ve been doing and this seems like a simple command that does a great job of achieving that
1 Like
Fwiw I’ve found a very cool flow, especially when i’m doing really large tasks, where i tell cursor agent (or claude code) to go through and understand what i’m trying to do and formulate plan WITHOUT ANY CODE CHANGES. I then have it write a full list of instructions as though it were explaining to an LLM what to do into a .txt file.
Then I just start various sessions with the agent saying “do tasks 1 and 2 from the insturction file” and then clear each time they finish. Works incredibly well and works with 3.7’s desire to rebuild the empire state building every time it finishes something
In claude code its possible to compact a conversation to reduce credit usage. Cursor only does this if the context size is large enough - it asks to summarise.
Users should be able to choose their compacting model and do so at any time. Ideally with a button.