Agent ignores Cursor Rules and Skill files despite explicit read instructions

[ucla98]

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

The AI agent consistently skips reading Cursor Rules (.cursor/rules/)
and Skill files even when:

• Rules are set as always_applied workspace rules
• Skills contain explicit “read this file first” instructions
• User explicitly says “read all skills and references before starting”

The agent says “I will read them” but does NOT actually call the Read
tool. It writes code based on assumptions, causing repeated errors.

Example: ADC channel was configured as CH4 (PC4/PC5) for multiple days.
main.h clearly defines ADC1_HVIN_V__Pin = GPIO_PIN_0 (PC0 = CH10).
Reading main.h would have revealed this in 5 seconds.
The agent never read it until the user pointed out the error.

This issue has recurred numerous times, and in each instance, the AI modified existing code without authorization, effectively resetting my project to square one. By the time I realized the AI was failing to follow instructions, a whole day had already passed. I would find the codebase in total shambles—the AI had spent the entire day injecting hard-coded snippets that completely violated my defined skill rules and overwriting my established code.

Discovering the cause of this issue only now, nearly two months into development, has been incredibly frustrating. I am beyond displeased. This failure has cost me invaluable time, and as a result, I have been unable to deliver the completed code to my client, long surpassing the promised deadline.

Steps to Reproduce

1. Create .cursor/rules/ with always_applied rule:
   “Read INDEX.md and all skill files before writing code”
2. Create skill files with explicit read-first instructions
3. Start a new chat session
4. Ask agent to implement a feature
5. Observe: agent writes code without calling Read tool
6. Agent makes errors that would be impossible if files were read
7. Point out the error → agent reads file → confirms the mistake
8. Repeat next session: same pattern occurs again

Expected Behavior

When Cursor Rules or Skill files instruct the agent to read specific
files before acting, the agent should actually call the Read tool
and verify file contents before writing any code.
A compliance mechanism or verification step is needed.

Operating System

Windows 10/11

Version Information

Version: 3.0.13 (system setup)
VSCode Version: 1.105.1
Commit: 48a15759f53cd5fc9b5c20936ad7d79847d914b0
Date: 2026-04-07T03:05:17.114Z
Layout: editor
Build Type: Stable
Release Track: Default
Electron: 39.8.1
Chromium: 142.0.7444.265
Node.js: 22.22.1
V8: 14.2.231.22-electron.0
OS: Windows_NT x64 10.0.26200

For AI issues: which model did you use?

claude-4.6-sonnet-medium (Cursor default agent model)

For AI issues: add Request ID with privacy disabled

7ca879e1-3351-4619-bf6d-feb5c01005ba

Additional Information

This issue repeats across multiple sessions.
The agent has workspace rules (.cursor/rules/) set as always_applied
that explicitly instruct reading INDEX.md and skill files before coding.
Despite this, the agent skips reading and writes code from assumptions.
The user must manually point out each error caused by unread files.
This has caused days of wasted debugging time on an STM32 firmware project.

Does this stop you from using Cursor

Sometimes - I can sometimes use Cursor

[Colin]

Thanks for the report!

Would you be able to share a minimal example repo (zipped or on GitHub) that shows this happening? Seeing your exact rule and file setup would really help us narrow it down.

As a reference, here’s a small reproducer we put together (attached). As you can see in the screenshot, the agent picks up the alwaysApply rule and reads the config file before writing any code. You can try it yourself by opening the project and prompting: “Add the ADC initialization function in main.c”

rules-repro.zip (2.9 KB)

One thing worth noting: if you have a rule that says “read all skill files before starting,” that can be unreliable since the agent has to decide which skills are relevant based on their descriptions. Instead, try making each skill’s description specific enough that the agent triggers it automatically for the right tasks (e.g. “Use when writing peripheral init functions” rather than relying on a blanket “read everything” rule).

If you’re seeing different behavior with a similar setup, that example would go a long way in helping us figure out what’s going wrong.