I have a rule that instructs the agent to always begin every reply with “test string”. The different agents always reply to “hello” with something like the following before reading the rules.
Hello! I can see you have a foobar project set up. Before I help you with anything specific, let me first read the core rules as instructed to ensure I follow all the guidelines for this workspace.
Then it reads the rules and responds again, this time following the instruction.
GPT-5 specifically, completely ignored the rules and I had to tell it to look at the rules. Then Cursor offered to save it as a memory but then GPT-5 replies and then reads the memory and then reads the rules and then replies again, this time following the instructions.
Expected Behavior
It should read rules before saying even a single word.
Operating System
Linux
Current Cursor Version (Menu → About Cursor → Copy)
Yes I just said “hello”. I’m testing to make sure my rules are recognized and followed and I thought this would be the best way to test.
I assumed for any input, even “hello” that in the background, cursor is telling the agent to read the rules first. I am new and learning how to use rules.
I removed the memory from earlier and then gave the agent a task to create a markdown file in the root of the project for me.
I figured out that previously the only reason it was even seeing the rule is because I had the rule file opened in a tab so it was added as context. Now with it closed all agents ignore it.
.cursor/rules/cursor-instructions.mdc
---
alwaysApply: true
---
# Cursor Instructions
Always apply all rules in LLM-Rules/core-rules.md
LLM-Rules/core-rules.md
# Core Rules
- Always begin every reply with: "Hello! I am Maximus Foobius, defender of the Foobar!\n"
