Haven’t used cursor but I assume it’s like other harnesses where the special sauce is in how it manages context, schedules sub agents, feeds them context, enforces standards, uses tools and skills etc. that make it better then just directly prompting the model.
For example you prompt opus directly with “refactor the auth flow” and it’s going to try and “one shot” it and produce the code from that prompt. Whereas a harness has instructions to say
Research the current implementation
Search the web for standards
Ask the user questions on how they want to do it
…
Claude code’s harness isn’t the best, but it’s still better than just promoting the model directly. One thing cursor has going for it is the vector embeddings of your codebase to make the harness more token efficient, but I still switched from cursor to managing parallel Claudes in cmux a few months ago and have had little reason to go back.
Haven’t used cursor but I assume it’s like other harnesses where the special sauce is in how it manages context, schedules sub agents, feeds them context, enforces standards, uses tools and skills etc. that make it better then just directly prompting the model.
For example you prompt opus directly with “refactor the auth flow” and it’s going to try and “one shot” it and produce the code from that prompt. Whereas a harness has instructions to say
Which produces way better results
Claude code’s harness isn’t the best, but it’s still better than just promoting the model directly. One thing cursor has going for it is the vector embeddings of your codebase to make the harness more token efficient, but I still switched from cursor to managing parallel Claudes in cmux a few months ago and have had little reason to go back.