Quantico

AI & Tech Creator | ex-Research Engineer @GoogleDeepMind
I build in public - the stack, the logs, the parts that break

[Quantico @web3_FBI]

"Agencies charge $5,000 for a portfolio site that looks this good. I built mine in 2 hours"

He's not exaggerating - and the clip below proves the pattern isn't a one-off.

Marissa shipped an app to 20,000 users for $230 total. The AI was the cheapest line: $51.

The build collapsed. A $10k level site is a Claude Code afternoon now. But her biggest bill was $142 of hosting - and it spiked because people actually showed up.

That's the lesson hiding in both: building a $10k looking thing is the cheap part now.

The cost - and the challenge - moved entirely to getting someone to use it.

Read the 2-hour walkthrough ↓ Just know the hard part starts after you ship.

[monokern @monokern]

https://t.co/qxOGBVt0UC

[Quantico @web3_FBI]

@misat0x Exactly why the app is built the way it is - people quit trackers by week two because logging is a chore. Remove the friction (just say what you ate) or
retention dies. It's the metric we watch hardest

[Quantico @web3_FBI]

Two people. One AI calorie app. [$14K]/mo.

My partner and I built a real app - Calorize, an AI calorie tracker - mostly by directing agents

The build almost cost more than it earns - until we stopped obsessing over the effort slider everyone's posting about and fixed the knob they skip: the model.

Effort = how hard it tries. Model = what it knows. We routed both per stage and the bill fell [$210 → $40]/wk for the same output. The split, the 5 prompts, the numbers ↓

[Quantico @web3_FBI]

https://t.co/iQd2jDlD71

[Quantico @web3_FBI]

Nice - same principle, different axis. You route by user knowledge level, we route by task stage, but it's the same move: match capability to what the moment actually needs, instead of one model for everything

Curious - do you split the verify step onto its own model too, or is it one model end to end?

[Quantico @web3_FBI]

@maxcryptoops Thanks

Honestly the margin came as much from kiling the build cost as from revenue - that's the part nobody posts about

[Quantico @web3_FBI]

Both matter, but you're pointing right at the part most people skip.

Swapping the model is the visible move. The workflow - which model + effort goes to which stage, plus a separate agent to verify - is where the real cost and quality live

The model is your ceiling. The routing is whether you reach it cheaply

[Quantico @web3_FBI]

@kulon077 Right - a bigger budget just raises the ceiling you waste against. Orchestration decides how much of it you actually spend.

Same model lineup, routed per stage, ran ~5x cheaper for us at the same output

[Quantico @web3_FBI]

@BzNav_R_Mont Well put. The visible lever gets all the posts; the second knob has no slider, so it gets ignored.

Routing beats cranking because it's judgment per stage, not a global setting

The founder/tooling parallel holds perfectly🦈

[Quantico @web3_FBI]

Every Claude Code clip: "effort just controls how deep it thinks"

Wrong framing - and it nearly cost me more than my app makes

Effort is how much work Claude does: files read, tests run, how far it pushes before checking in. "Think harder" is the small part.

And every clip assumes one model. That's the second knob.
Effort = how hard it tries.
Model = what it knows.

Building Calorize with two people, the fix wasn't the slider. It was routing both knobs per stage. Bill dropped [$210 → $40]/wk, same output.

Two knobs. Most people only touch one.

[Quantico @web3_FBI]

https://t.co/iQd2jDlD71

[Quantico @web3_FBI]

One test per stage: could I write the instructions exactly?

Then small model - edits, migrations, wiring. If it needs judgment (ambiguous logic, edge cases, anything a user trusts a number from) - bigger model, higher effort

In Calorize the meal parser was the one place I paid up. Everything around it ran cheap

[Quantico @web3_FBI]

@aiseomastery Yeah, you're right

[Quantico @web3_FBI]

"Agencies charge $5,000 for a portfolio site that looks this good. I built mine in 2 hours"

He's not exaggerating - and the clip below proves the pattern isn't a one-off.

Marissa shipped an app to 20,000 users for $230 total. The AI was the cheapest line: $51.

The build collapsed. A $10k level site is a Claude Code afternoon now. But her biggest bill was $142 of hosting - and it spiked because people actually showed up.

That's the lesson hiding in both: building a $10k looking thing is the cheap part now.

The cost - and the challenge - moved entirely to getting someone to use it.

Read the 2-hour walkthrough ↓ Just know the hard part starts after you ship.

[monokern @monokern]

https://t.co/qxOGBVt0UC

[Quantico @web3_FBI]

https://t.co/iQd2jDlD71

[Quantico @web3_FBI]

https://t.co/iQd2jDlD71

[Quantico @web3_FBI]

@Caarat1 Yes, yes, attention is everything to us. It's easy to lose and hard to regain

[Quantico @web3_FBI]

@crytonbuton Yes, agree 🙂

[Quantico @web3_FBI]

@CSLeSVyDU7XyKpM bot???

[Quantico @web3_FBI]

@Timur_Yessenov @catmanyau "Paying per unit of ambiguity" - that's the whole thing in five words
Everyone optimizes the model; the leverage was always in the brief

[Quantico @web3_FBI]

Exactly -and it quietly raises the bar to raise at all. When a working build is $51, a demo stops being a moat; anyone can vibe-code one.
So seed isn't funding "can they build it" anymore - it's funding distribution, speed, and a team that can out-execute the 50 others who shipped the same thing this week. The product was never the hard part. It just became obvious.

[Quantico @web3_FBI]

@PatriaHeapin Merci - glad it landed.
The "build is cheap, distribution is the game" shift is going to reorganize way more than people expect🤝

[Quantico @web3_FBI]

@misat0x Yes, agree

[Quantico @web3_FBI]

Fair - you're right it's moved to credits/tokens.

But your core point is the one that matters: iteration count conflates two things-real complexity vs how much you flailed. 600 messages could be a hard app or a vague brief

The lever nobody names is spec quality - the tighter your first prompt, the fewer revisions. You're basically paying per unit of ambiguity.