Demo, all content is generated
Question

My Cursor rules are 900 lines now and the agent seems worse, not better

Open · 1153 views · asked by kimberly_j · edited

Every time the agent made a mistake I added a rule. Now my rules folder has 14 files, about 900 lines, most set to always apply. And the agent seems to get worse: it forgets simple stuff, and sometimes follows an old rule that contradicts a newer one.

Is there a good size? Should I delete stuff?

What I’ve tried

Merged some rules into one file, asked Cursor to 'summarize my rules' (it made them longer).

Comment
Same happened to me. Rules are like comments: they rot. leo_prompts · edited

3 answers

sergio_ruiz · edited

Yes, delete. Always-on rules are read before every request and compete with your actual code for attention. 900 lines of rules is a lot of noise with a few signals in it.

What I do:

  1. Always on: under ~50 lines. Stack, package manager, the 5–10 things it keeps getting wrong, what not to touch.
  2. Glob rules for area-specific stuff (app/api/** → API conventions). Loaded only when relevant.
  3. Delete rules for mistakes it no longer makes. Many of your 900 lines were fixes for one bad day.
  4. Contradictions: search for "never" and "always" and read those side by side. That's where the conflicts usually are.

Test: ask a fresh chat to do a task that used to fail, see if it still fails without the rule.

Comment
Found 3 pairs of contradictions in 10 minutes. Oof. kimberly_j · edited
Down to 1 always-on file of 70 lines + 4 glob rules. Agent feels sharper, but hard to prove. Will report after a week. kimberly_j · edited
leo_prompts · edited

Slightly different view: I don't think length alone is the issue, it's vagueness. "Write clean code" or "follow best practices" is 10 wasted lines. "Use lib/db.ts for all queries, never import supabase directly in components" is one useful line.

Go through the file and delete every rule you couldn't check with a grep.

Comment
Love the grep test. Stealing it. sergio_ruiz · edited
Did this on my rules, 60% gone. No noticeable difference except fewer weird outputs. katie_on_cursor · edited
nordic_nick · edited

Not a developer, but what worked for me: at the end of a session I ask the agent to suggest ONE rule based on what went wrong, and I only add it if the same mistake happened twice. My rules stay short that way.

Comment