How I Use Cursor (Now with GPT-5)

I didn’t expect to write this today, but GPT-5 dropped, and Cursor is giving paying users free credits for launch week.

That’s basically the equivalent of finding a $20 bill on your desk (and then some, because it seems they gave me extra credits). Let’s talk how I use Cursor day-to-day.

The magic comes from shaping context, often adding as much as possible_._ The context window is massive compared to a year ago, so use it! Over time, I’ve built a small set of practices that make the AI supercharge my workflow. Hopefully I’ll discover more tricks, now that GPT-5 is out.

1. My default.mdc Is an Index File, Not a Dump

One early mistake I made: cramming everything into default.mdc.

That’s like what I used to do as a high school student: carry all textbooks everywhere instead of looking at the classes for the day and brining books that I actually needed.

Now my default.mdc looks more like this:

Please follow best practices, depending on your task: 
- `best-practices-code-style.mdc` for code style 
- `best-practices-cursor-agent-mode.mdc` for Cursor Agent mode 
- `best-practices-data-access.mdc` for data access conventions 
- ... 

After a task, succinctly list which Cursor rules you used.

The “rules” aren’t rules per se. They’re just guidelines, so the Agent doesn’t get lost. When we get lost as humans, we look for a guide — it’s the same here. By keeping the index slim, I’m trying to avoid context and usage limit headaches and the AI can grab what it needs, when it needs it.

I ask the model to list Cursor rules it uses, because I’m currently trying to see if it helps me optimise the look ups.

2. Best Practices Driven

Best practices aren’t rules. They’re what you reach for when you’re not sure. For me, they cover:

• Code Style — Comments explain why, not just what. Imports grouped. Names that actually mean something.

• Data Access — Always through a DAO tier.

• Error Handling — Fail loud where it matters, dead code is better than limping code, validate inputs, don’t let async errors disappear.

• React — Functional components, hooks, and typed props.

• Testing — it() for behaviour (instead of less fluent test()), formatting the describe() statement.

These files are living. Every time I trip over something, I think about adding it to the best practice docs.

3. AI Dev Tasks: The Not-So-Secret Secret Weapon

Cursor’s Agent mode is great, but AI Dev Tasks have become the go-to workflow upgrade. If you haven’t seen Ryan Carson’s repo: snarktank/ai-dev-tasks. Have a look.

I use three rules:

• create_prd — It’s like having a PM who asks the really good clarifying questions. The questions that require executive-level input.

• generate_tasks — Turns the PRD into a concrete task list. Tactical.

• process_tasks — With tasks broken down sufficiently, execute them.

It’s not just about prompt optimisation — it’s about context optimisation. With these prompts, the AI knows where we’re going and why.

4. Brand Guidelines Aren’t Just for Marketers, Designers

The other file that is underrated: brand guidelines.

I do a lot of front-end, landing pages, new feature work and copy is often “TBD.” Instead of stalling, I feed the AI my brand voice doc and get solid placeholders instantly.

Mine covers everything from:

• Tone of Voice — Sample phrases, dos and don’ts.

• Color Palette + Typography — Get the vibe right.

• Layout Principles — Mobile-first, modular components.

• Key Benefits + Differentiators — So even initial copy feels on-brand.

It means I can keep developing while waiting for final copy. When product managers or designers review your work, they’ll see “good enough” copy to sign it off.

Closing Thought

GPT-5 in Cursor is fun but it still needs structure. The real unlock is to treat AI like a junior dev who can out-type me (times 1,000,000) but still needs the right guardrails.

If you’re experimenting with AI in your workflow, I’d love to hear how it looks like.

I write more like this at blog.mariohayashi.com, and you can find me on Twitter: @logicalcy.