Cursor is currently brimming with features and models. It's quite hard to identify what the optimal way of using it is.

I am a somwhat experienced developer who has taken some time off work, and I'm currently using Cursor Ultra 8+ hours every day to build out my personal project.

I wanted to share my findings as someone who is free to experiment as much as they want:

Composer is King for Edits

Composer is king for edits. It's fast, accurate, and relatively cheap.

Composer is not king for good decision making, which is why you need to use other models to set up sufficient context around Composer to implement changes properly, e.g., with plan mode to create prompts.

Plan Mode vs. Ask Mode

"Plan mode" is nice in theory—but quite buggy and inconsistent in practice. It is also a bad tool for interrogating a plan—you are taken out of the conversational back-and-forth that LLMs are optimized for, especially as you refine a feature or change you want to be making.

"Ask mode" is king here—any time I want to make a complex change, I start in ask mode with a smarter model. Gemini Pro 3 is great when it works, but it doesn't usually work for complex tasks (unrecoverable loop errors, capacity issues, throttling, etc.). GPT-5.1 has been my go-to, striking a very good balance between speed, quality, and availability.

My Workflow

My workflow is as follows:

I start in ask mode by discussing the feature/change—which files it is likely to touch, what I want, how I think I want it. And then, crucially, I append this to the end of my prompt:

With all this [previously explained] in mind, can you evaluate what needs to be done, identify the changes that need to be made, and whether anything needs to be reworked/rearchitected to ensure a clean, simple, and elegant design?

That last bit is relevant to me and my project, where I can afford to be constantly re-architecting my codebase. Where applicable, it's still a very useful imperative, as it discourages models from designing shortcuts and building on top of poor decisions made elsewhere.

In ask mode, GPT-5.1 will come back with code examples, explanations, and rationale for proposed changes. You have a great opportunity to dive deep into the proposed solution beyond a high-level .md prompt as proposed by the "Plan mode," and you will be able to steer the model towards a proper solution with less effort. When you switch to making edits, the next model will not have to deal with ambiguity around what needs to be done.

When happy with the proposed changes, I switch to Agent mode with Composer, and simply tell it—"Let's implement." 8 times out of 10 I've found the results to be exactly what I need.