It's not about the spending of AI tokens, it's about the potential of prototyping.
I like to build things in my free time.
Not necessarily products. Not necessarily monetization ideas. Just things I want to exist, things that benefit myself, or maybe others.
And I use AI a lot for it.
Because it's no longer about writing the best possible code upfront.
It's about having a prototype ready, fast, in some kind of MVP state. Then iterating.
Building with AI + your own vision
I recently redid my website for my weekend projects:
I have to say, when you combine your own visual idea with AI, it gets really close to what you actually had in mind.
This is not a promotional post.
It's more about showcasing how empowered we are now to just… build.
Small note though:
Yes, humans are still accountable for the code.
I don't expect AI to write production-ready systems. Not without iteration, validation, and actually understanding what’s going on.
What I’ve been building
I’ve always had multiple projects going on. Games, tools, apps… always something upcoming.
🎮 Games
Main idea:
playable in the browser
no login
no tracking
just open and play
I care more about the experience than anything else.
I started with Warlocks, inspired by the Warcraft 3 days.
Pure chaos: pushing, dodging fireballs, shrinking safe zone.
Once I hit stable 60 FPS, I kept pushing:
bigger maps
more players
more interaction
Then came something inspired by Pudge Wars-style gameplay:
Mortar Mayhem
16 players, 8v8, blind mortar shots across the map, fog-of-war feeling, upgrades based on performance.
Then:
Pirate Ships deathmatch, 9 players, real-time aiming and prediction
Tank Blitz, same idea, but with powerups, turret alignment, charged shots affecting accuracy
At some point it became a pattern:
Boats. Tanks. Submarines. Planes. Bicycles.
Characters with bows. Pistols on horses. Whatever.
The constraint is always the same:
60 FPS in the browser.
🛠️ Tools
These usually come from real needs.
llm-schema-guard:
Strict schema validation + enforcing correct shapes + observability
commerce-orchestrator-api:
Moving towards agentic commerce, a layer in front of backend systems so agents don’t integrate directly with everything
Apex-Edge:
One problem I’ve seen everywhere:
What happens when the internet goes down? Or AWS/GCP?
Retail just… stops.
Idea:
local hub per store or region
process transactions locally
sync later
don’t push all complexity into POS devices
📱 Apps
Sometimes it’s just solving annoyances.
the-fuse
I was streaming and didn’t always have people to play with → so I built something to match creators for collaborations
what-i-paid
I got tired of not tracking prices across stores
→ take a picture of a receipt
→ AI extracts everything
→ track, manage, share
Free, simple, low friction.
And a lot more…
There are a lot of other ideas in progress.
Some go more into AI:
local assistants
multi-device control
something like a lightweight “brain” coordinating things
Some go bigger into games.
What I learned
Adapt to AI and iterate fast.
That’s it.
Prototyping is cheap now.
You don’t need to sit on ideas for weeks anymore. You can build something in hours or days, see if it works, and move on or improve it.
Just don’t forget:
If you ship it, you own it.
AI doesn’t change that.
(Quick note: I only linked the hub page because this isn’t a promo post. I just wanted to share the mindset and the joy of building fast with AI. The more we all build and iterate, the better it gets for everyone.)
Top comments (15)
The reframe from "token costs" to "prototyping velocity" is the right lens. The calculation isn't cost-per-token, it's cost-per-prototype, and what AI does to that number is dramatic. A prototype that would have taken a weekend now takes an evening, which means you can run more experiments, kill bad ideas faster, and arrive at something worth investing in with much less sunk cost.
The "things I want to exist" motivation is also underappreciated in how it produces different outcomes than "building for a market." When you're building something you personally want, you have a much cleaner signal on whether it's working — does it do the thing you needed? That's a harder question to answer honestly when you're building for a hypothetical user.
The iteration model you're describing (MVP fast, then iterate on what you actually discover vs what you assumed) is the practical benefit AI unlocks most reliably — not better code, but faster feedback loops.
Thanks so much, this is one of the most insightful comments I've gotten on the article. You nailed exactly what I was trying to get across.
Totally agree on the "things I want to exist" motivation too. When I'm building for myself, the feedback loop is brutally honest, either it scratches the itch or it doesn't. No hand-waving about "user personas" or "maybe they'll like it." It's been a game-changer for actually shipping stuff.
Appreciate you taking the time to write this out. Comments like yours make writing these pieces worth it. Have you noticed similar shifts in your own prototyping workflow with AI?
This really resonates. The shift from "write perfect code" to "get a working prototype fast and iterate" has been a game-changer for me too.
I've been running a similar workflow on a large-scale Astro site — using AI to generate content for thousands of pages, then iterating based on real data from Search Console and analytics. The feedback loop is so much tighter now. What used to take weeks of planning and spec-writing turns into a weekend prototype you can actually test against real users.
Your point about accountability is important though. I've seen AI-generated code that looks clean but has subtle data issues — things like wrong calculations that pass a quick glance but fail when you look closer. Iteration has to include validation, not just feature velocity.
Curious about the 60 FPS constraint for browser games — are you using canvas or WebGL for rendering?
Thanks! Glad it resonates.
For the 60 FPS constraint: All my games are 3D. I mainly use Three.js or Phaser.js with backend-authoritative rendering, and I try to disable frontend re-renders / state manipulation entirely to keep consistent performance.
I've actually recently made the sync part public, here: github.com/AncientiCe/realtime-syn...
This is the magic that syncs at 60 fps.
Of course this really fits my type of projects, might not fit yours.
Oh nice, 60 fps sync is no joke — that's way beyond what most prototyping setups handle. I'll definitely check out the repo. My projects are more on the static site / data pipeline side (thousands of pages generated from structured data), so real-time sync isn't my core use case, but I can see how that changes the prototyping feedback loop completely. Appreciate you sharing it!
Love this, your “prototype first, own what you ship” mindset feels refreshingly honest, and I think a lot more builders needed to hear that today.
nice
💯
token costs come up in every retro. but nobody asks how many ideas died before they could be validated because prototyping was too slow. the real leverage is earlier in the loop.
Exactly.
Token costs show up in every retro because they're easy to measure. But the silent killer is all the ideas that never even got prototyped because the friction was too high. AI moves the bottleneck earlier in the loop, and that's where the real leverage hides.
Appreciate you putting it so cleanly
yeah the bottleneck shift is the part that is hard to explain until you have actually lived it. when prototyping is cheap the question stops being can we afford to try this and becomes why didnt we try this sooner
Yup, same here, in a different way
Best wishes to your projects!
ghgkjoiukj