Agent ignores skill

Support Bug Reports
,
1

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

I experiment in Nightly with a skill below and Agent behaves as usual, rather than following the instructions from the Skill:

---
name: Techlead
description: How to be a Senior AI
---

# My Skill

You are the tech lead. Execute assigned tasks via sub-agents.

## Instructions

### Process

1. Cluster tasks by available agents.
2. Execute clusters from the most complex to the simplest.
3. After each batch, run verifier.
4. If verifier found an issue:
    1. if the problem is in criteria/lack of context — refine the prompt and send it back to the same agent;
    2. if the problem is architectural/complexity is higher — escalate to the next level.
5. Tech lead report: the order of agents invoked + the standard report by form.

### Constraints and recommendations

- CI is single-threaded — meaning only one sub-agent can work at a time.
- If applicable: in the prompt for a sub-agent, quote the TODO items you assigned it right away — so it doesn’t have to load the whole thing into context.
- If an agent’s work was interrupted by any error, run Verifier to find out which part of the work it managed to complete.
- With each new invocation, a sub-agent has zero memory. Context is not preserved between calls.
- If manual testing is required after completing a task (play the game), then:
    - either pause work on this task and note in the report that a manual check is required; describe in detail what exactly the Tester needs to do.
    - or stop work completely and request a check, if it’s an important blocking change.
- Cost of sub-agents: Deadlock Breaker >> Senior SWE > Middle+ SWE >> Middle SWE > Junior SWE
    - Try to invoke sub-agents economically, but keep quality standards high.
    - For documentation fixes, invoke Middle SWE, or make the edits yourself.
    - For large documentation changes, invoke Deadlock Breaker.

If the same text is pasted directly into the prompt, the agent works as intended.

Steps to Reproduce

  1. Set up subagents.
  2. Create a skill that commands their use.
  3. Give the agent a task, calling the skill at the end.
  4. The agent will behave as usual, not as specified in the skill.

Operating System

Windows 10/11

Current Cursor Version (Menu → About Cursor → Copy)

Version: 2.4.0-pre.36.patch.0 (system setup)
VSCode Version: 1.105.1
Commit: 5fdf39db168c663448e157fd8703e614773db640
Date: 2026-01-15T08:25:51.483Z
Build Type: Stable
Release Track: Nightly
Electron: 37.7.0
Chromium: 138.0.7204.251
Node.js: 22.20.0
V8: 13.8.258.32-electron.0
OS: Windows_NT x64 10.0.22631

For AI issues: which model did you use?

GPT-5.2 XHigh

Does this stop you from using Cursor

Sometimes - I can sometimes use Cursor

1 Like
(Continuously Updated) My Real-Time Review of Cursor Subagents
What is the difference between skills and rules?
3

Hey @Artemonim,

The Agent Skills feature is in beta (source).

Workaround: Use Rules instead - create a .cursor/rules/techlead.mdc file with your instructions and set the apply mode to “Always” or “Agent Requested” (manual). This should give you the behavior you’re looking for until Skills is fully released.

1 Like
4

What’s the point of Skills anyway if we already have Rules that the agent follows better?

1 Like
5

always applied rules eat into your context window, agent requested ones ofthen just dont work (or you dont know if they do)

skills are loaded too but only the frontmatter (saving context window space) and the agent can decide to read the rest if the frontmatter seems interesting enough

best usecase is actually manually activating them with slash commands

OR

you write a custom command/skill that references other skills, so youre kinda able to build a chained workflow

for me i have an exploration agent/workflow that is being called by my “create 37 highly informed and ultra detailed todos for … ) which uses the exploration, then todos, then critique steps every 5-10 todos, which then verifies the prior completed todos and updates the remianing ones if necessary (works best iwth Gemini 3, 2nd best with all Claude models, worst with GPT 5 and 5.2 as the oAI models stop at every second todo )

6

Yes, this should be a major point of discussions: Rules, Commands, Subagents, Skills and other Taxonomy Headaches