A single AI agent hitting an API can burn $300+ per day.

That’s close to $100K per year. Per agent.

Jason and Chamath (“All In” podcast) were breaking this down recently and it hit close to home because I’m watching it happen in real time.

We’re no longer managing flat SaaS subscriptions.
We’re managing variable compute. And variable compute compounds fast when nobody is watching the meter.

Right now, in most startups, nobody is.

Ask a founder, “What’s your monthly token budget per developer?” → Most don’t know.

Ask, “Who’s burning the most compute?” → Silence.

It’s like giving every engineer an uncapped corporate card and hoping the invoice doesn’t hurt too much.

And here’s the thing nobody talks about:

When a developer’s token usage starts costing as much as a junior engineer’s salary, that spend needs to produce leverage. Real output and acceleration, not just cleaner code comments.

As a result, the scary part isn’t the cost, it’s the invisibility.

Headcount shows up immediately on your P&L.
Token burn hides quietly in dashboards nobody checks until the invoice lands. And suddenly you’ve burned weeks of runway without realizing it.

We figured out how to use AI.

Now we need to figure out how to manage it.

Because very soon, your token bill might start competing with payroll.

Rork just dropped Rork Max, powered by Opus 4.6, and Tech Twitter went wild.

People are throwing random mobile app ideas at it and it’s almost one-shotting full native iOS apps. Real screens. Real flows. Surprisingly clean architecture for a first pass.

The bar just moved again.

You now “just” need a strong idea and a clear spec. Tools like this can generate a serious MVP in minutes.

The catch? Opus 4.6 isn’t cheap. It can burn through credits frighteningly fast.

But if you’re building something real, not just playing around, the leverage is insane. One focused session can replace weeks of boilerplate.

The game is shifting from “can you code?” to “can you direct the agent well?”

This week I finally experimented with Skills inside Cursor.

Skills are basically pre-configured instruction layers that make your agents operate in a specific, opinionated, precise way. Instead of prompting from scratch every time, you give the agent a structured “mode of thinking.”

The Skill I tested is called Vercel React Best Practices.
It embeds best practices for writing clean, scalable, and fast React and Next.js code.

Here’s what I did.

First, I installed it via the integrated terminal:

This creates the Skill file inside Cursor. After that, you can call it with the slash command: /vercel-react-best-practices

In my case, I asked the agent to review the pages inside the dashboard section of my project and analyze them against the defined best practices.

What happened next was powerful.

It pinpointed the issues, ranked them by importance, and then fixed them. Structured refactors. Cleaner components. Better patterns.

This is how you turn vibe-coded code into production-ready code.

You still move fast.
But now you add guardrails.

And this is just one Skill.

There are many more to experiment with. Cursor even launched a marketplace where you can integrate with popular dev tools like Stripe, Supabase, Linear, Slack and others. You can connect their official MCPs and Skills directly, without messy setup.

The bigger idea here is this:

Agents are good.
Agents with context and structured constraints are elite.

If you are building seriously, start thinking in terms of reusable Skill layers, not just prompts.

PS: I’m working on practical guides and a small live cohort to help non-developers actually ship real apps with AI (not just learn theory).

I’ll be opening this to a small group first,
join the waitlist here → https://tally.so/r/gDeLrM

Hope you enjoyed this aiOS App Builders edition!
💡 Got an AI tool for us to check out or collaborate?
Send us a message and let us know!

Was this edition forwarded to you? Sign up here

Filippo