I seriously can’t tell if you’re seriously or not, so I’m just gonna answer seriously:
1 Like
I could tell you to ask it about writing a rule so it doesn’t ignore it all, but…
Whenever something you don’t want happens, stop the generation and do this:
Why did you do X like Y, specify all the rules and instructions that may have generated this decision.
You will most likely detect a conflict, an oversight, lax rules, or specific instruction that induces Y in X, or several.
Go with the vibe!
For example, upon encountering the work of others in the profession of CONTEXT ENGINEERING, I would take his link and do the following:
Read @link and add any rule that is compatible and useful to our cursorrules while adapting to the general tone, restrictions and language/frameworks specified in our project.
Thus, anything interesting will be added and refined into the project’s context. This is highly dependent on the model used at the time of the request, using a different model to do this, or to use the result will not be optimal. If you use different models, rewrite the rules for each or import separately.
I type zero lines of code or instructions. Everything in the project is vibe engineered using the chat window.
Agent Compass works for all models. One by one rule was added after working with Claude 4 and Gemini 2.5 Pro, but they still work for all models.
Although I should add in the Readme that user can add their own style rules in section four.
What is your opinion on the XML > JSON > Markdown model rules respectfulness?
I have been “programming” various things in various domains for over 30 years.
I have written 0 (ZERO) lines of actual code in the last 12 months.
I have been generating useful applications, tools, documentation and linux system configuration/customization by just asking the AI run sudo commands.
I am almost 50X more productive and 50X more exhausted because my creativity is drained by the machine faster than it can recharge.
Not entirely. The quality of the output is based on the quality of the input. If you are casual about it, so will be the model. If you are super abstract, so will the model result. Cursor provides a foundational ruleset to kickstart your progress. If you encounter serious issues, the problem is usually at the model level. It doesn’t have enough coherence to generate familiar output.
Try to start simple and plan each small step, like building with legos.
If you notice the agent generating new files instead of adapting existing files to new functions, then yes, definitely you are to blame.
If you start from an existing project, you need to tell the agent to document itself. But otherwise, the only rule that really matters is:
Just tell the agent what to do.
1 Like
While writing this post, I clocked another 1k LoC ontop of the 15k.
The agent was given the PRD and:
Assign a version increment to this feature.
Do not add extra columns after score.
Reformat the table without borders, you may use colors.
Implement the features as planned.
Run and test the output of the implementation.
Update md files upon completion as needed.
I have paid $20 for about $60 worth of value. Until next month I will pay about $2-$3 per hour of use, if I find it’s worth it to spend that. You can rack up to 500k a year easy on Cursor if you really want, look at the screenshots from this thread. Don’t use Auto mode for coding serious things, try other tools/services/plugins or pay for usage.
1 Like
Almost the same usage. Ultra plan is not sufficient to cover this.
