Stack — Diana Seo

Stack — how I build

tools · hacks · takes

I build like an operator, not a hobbyist. The AI tools aren't novelties — they're an operating system for shipping. The edge was never access to the model. It's knowing what to ask, what to throw away, and how fast to move.

operator console — build & ship the Oriya normalizer, in one night · illustrative

Manus

›

GPT

›

Claude Code

›

Replit

›

Evals

›

Vercel

raw readout →one fair 0–100 index

Ourareadiness 82—

Whooprecovery 44%—

Garminbody battery 55—

Appleno native score—

Same night, four devices, four scales that don't compare — 82 vs 44% vs 55 vs nothing at all. Give each one a handicap and they land within a few points of each other: now a squad on mixed hardware competes fairly — and it held across 50 synthetic edge cases before it ever shipped. That normalizer is Oriya.

01AI · building

Claude Codeagentic coding

My primary pair on real codebases. I scope tightly, let it hold the architecture, and review every diff like a staff engineer. The model isn't the moat — knowing what to ask is.

Replitprototype → deploy

Zero to running app in one tab. Where an idea gets tested before it earns its own repo.

Vercelship the frontend

Preview deploy per branch. “Show me” becomes a live URL in minutes, not a meeting.

Manusautonomous runs

I queue the grunt work — competitor teardowns, data pulls, first-draft ops — then audit the output. Delegation, not magic.

GPTreasoning · adversary

My scratchpad and sparring partner. I make it argue against my own thesis before I commit code or capital.

Geminilong context · multimodal

Drop in a whole deck, transcript, or spec and interrogate it as one object. Breadth where the others go deep.

02Productivity · SaaS hacks

Build the throwaway first. A 30-minute prototype kills more bad ideas than a three-hour debate.

Prompts are infrastructure. Version them, reuse them, treat the good ones like code — not clever one-offs.

Automate the second time, never the first. The first pass teaches you where the edges are.

One triage queue. Route every tool's pings into a single surface — context-switching is the tax nobody prices.

Default to the boring stack. Novelty is a cost; pay it only where it's the actual product.

03Migrations · the switch

Namecheap → Cloudflaredns · edge

Left the domains at Namecheap out of inertia until a record change took hours to propagate instead of seconds. Cloudflare runs the fastest anycast resolver on the internet, registers domains at wholesale cost with zero upsells, and hands you a WAF, unmetered DDoS mitigation, and edge caching on the free tier. The pitch that finally moved me wasn't a feature — it was that I stopped thinking about DNS entirely.

Netlify / raw AWS → Verceldeploy

Netlify was the gateway drug; raw AWS was the tax I paid to feel serious. Vercel won on one mechanic: a preview URL per branch turns “show me” from a meeting into a link. Framework-defined infrastructure means I stopped hand-writing the deploy, and because every build is immutable and promotion is just an alias swap, rollback is instant — no rebuild, no downtime.

Postgres read-replica → Snowflakeanalytics

Analytics rode a groaning Postgres read-replica until the dashboards started timing out and every heavy query threatened prod. Snowflake's unlock is boring and total: storage and compute are decoupled, so a brutal backfill spins up its own warehouse, never touches the app, and you pay for the seconds it runs and switch it off. Separation of concerns, billed by the second.

EC2 GPU babysitting → Modalai infra

A newer one I genuinely enjoy. For anything that needs a GPU for ninety seconds, Modal killed the old loop — spin up an instance, forget to kill it, find a $400 bill next week. You decorate a Python function, it runs on their fleet, you pay for the execution and nothing idles. Serverless the way the word was supposed to mean.

04Stack insights

The bottleneck moved from writing code to specifying intent. Taste and judgment are the scarce inputs now.

Most “AI products” are a system prompt in a trench coat. Defensibility lives in the data loop and the distribution — not the model call.

Speed is strategy, not a personality trait. Whoever learns fastest in-market wins; tooling just compresses the loop.

Owning the workflow beats owning the model. The model is rented; the workflow, the data, and the trust are yours.