Cursor Agent not following RULES

Support Bug Reports
1

Has anyone noticed that the Cursor Agent is no longer using/referencing any of your RULES? I’m getting slop, and when I show the Agent the rule, it didn’t apply; it can confirm it is not using them at all. (MDC version of RULES)

Version: 2.3.41
VSCode Version: 1.105.1
Commit: 2ca326e0d1ce10956aea33d54c0e2d8c13c58a30
Date: 2026-01-16T19:14:00.150Z
Electron: 37.7.0
Chromium: 138.0.7204.251
Node.js: 22.20.0
V8: 13.8.258.32-electron.0
OS: Darwin arm64 25.2.0

3

Hey, thanks for the report. This is a known issue. The agent sometimes ignores rules from MDC files. The team is aware.

To pass details to engineering, we need a couple things:

  • The Request ID from the chat where the rules were ignored: chat menu (three dots on the right) > Copy Request ID
  • Which mode are you using, Ask or Agent?
  • Which model, Auto or a specific one?
  • The path and approximate contents of the MDC files (if you can share)

Related threads: Cursor ignoring .cursorrules Cursorrules getting ignored

Can you share the Request ID and the details?

4

Yes. Please see below:

  • c46a6c0c-0223-4d3f-9614-f1014c91433e
  • Mode: Agent
  • Model: Optus 4.5 and GPT-5.2-Fast
  • Project Root Path: .cursor/rules/xxxx.mdc
5

What is interesting is that if I create a new agent and ask it, “What are your RULES?” It lists all the rules by Always-Applied Workspace Rules and Agent-Requested Rules. :stuck_out_tongue:

Request ID: 7272adde-955e-4979-8e44-b0bdba865867

6

Perfect, thanks for the Request IDs and details.

Interesting point that the Agent can see the rules when asked directly, but doesn’t apply them while working. That’s useful info for the team.

If you can, please share the full contents of your .cursor/rules/xxxx.mdc file (or at least the main structure). It’ll help with debugging.

9

Hi @deanrie - Sorry for my delayed response… I’m based in Sydney, Australia…

CleanShot 2026-01-23 at 09.08.14@2x
CleanShot 2026-01-23 at 09.08.14@2x458×840 65.9 KB

I used the .txt extension so I can upload the file here. :stuck_out_tongue:

docs-style.txt (2.7 KB)

teammate-mindset.txt (3.5 KB)

research-rigour.txt (8.8 KB)

10

Thanks for the files and the details.

The issue is that Privacy mode blocks access to the Request ID for debugging.

While the team is digging into this, here’s what you can try:

What you can try:

  1. Check Cursor Settings > Rules. Do your .mdc files show up as “Always Applied”?
  2. Try temporarily turning off Privacy mode (Cursor Settings > Privacy), run a test Agent request, and share the new Request ID. This will give the team more data.

What’s confirmed:

  • You do have alwaysApply: true in the frontmatter
  • The Agent can see the rules when you ask directly, but it doesn’t apply them during the task

Let me know what you find if you try the steps above. It’ll help with debugging.

11

Hi,

I confirmed that Cursor IDE shows these rules as ALWAYS. :slight_smile:

I temporarily turned off Privacy (Team account) and started a new thread.

  • Auto: 0983c900-0776-4363-81db-b8c773cd13df
  • Optus 4.5: d7477088-168d-4fae-a0e7-543c5be1d113
  • Gemini 3 Pro: b78a9e76-e427-46ee-9c32-2a103a0a20b0
  • GPT 5.2 High (did a better job): 1366b1b5-b816-4044-ac35-07fb15d450be

T

12

I’m currently specifying in AGENTS.md that she needs to read my rules structure and tell me which rules she read.I’m currently specifying in AGENTS.md that she needs to read my rules structure and tell me which rules she read.

14

We checked your Request IDs, the rules are loaded into the context, and technically everything works correctly.

There are two points here:

  1. Irrelevant rules: for the task “Review all README.MD files”, rules like python-development, database-connection, and qa-quality just don’t apply. The agent can see them, but won’t use them because they’re not related to the task. This is expected behavior.

  2. Relevant rules (research-rigour, teammate-mindset): the agent admitted it skipped them and fell into the “Helpful Assistant” trap. This is an LLM limitation. Rules in context don’t guarantee 100% adherence. These models are probabilistic.

What you can try:

  • Make critical rules shorter and more specific
  • Add an explicit “ALWAYS follow for ANY task including documentation”
  • For important tasks, remind it in the prompt which rules to apply

So it’s not a bug in rule loading, but I’ll pass the adherence feedback to the team.