I see a lot of this idea that “more tokens, maxxed models, and more overnight grinding equals better output”. I disagree.

There’s a lot of noise lately from influencers bragging about the thousands they’re spending every month in tokens to write code. Folks setting dozens of high-end (expensive) models loose on a problem overnight or over days. I even saw one guy spend $1000+ to make a new app, then validated the results mechanically by asking another prompt to verify the features were complete. He didn’t even look at the work.

I Do Things Differently

My rig is simple. I use Claude Code running the Haiku 4.5 model. It’s fast, cheap, reasonably capable, and doesn’t spend a lot of time overthinking. I treat it like a junior dev who codes like the wind, uses Google a lot, has a little too much confidence, and makes mistakes.

Then I pair program with it. If I need more capacity, I’ll open another terminal and pair with another instance working on a different part of the problem. At most I’ll have 3 running, but I’m only focused on 1 at a time.

“But how can you be 10x?” I hear the cool kids ask. “Where’s the tokens, bro? Why are you spending time working with the stupid things instead of influencing others while your token-eating machine whirrs in the background?”

Great Results, Low Costs

I’m happy with it at least. In the past 5 months I’ve shipped:

• a new language and runtime with a robust ecosystem and full docs (duso.rocks)
• a new AI-based iOS app with its own server (written twice, once in Go and then again, leaner, in Duso, for reasons) (arland.ai)
• 4 new websites (3 static-built with Duso, 1 live)
• 2 internal AI workflow tools (1 with a full interactive web-based GUI)

All working well. My cost? Two months at $20/mo and 3 months running $100/mo. That’s $340 total for 5 months of shipping real things. Not thousands of dollars per month, but dozens.

Applying Good Practice

This is about applying decades of good practice to a new tool, not trying to avoid work. When the dust settles, AI coding won’t look much different than other tools like 3D printers, CNC machines, even Photoshop. They all allow more people to participate, but to do well they’ll need to learn their craft. Highly skilled makers will be able to leverage the hell out of it, efficiently, and the rest will fizzle out.