Over the past two days I built three things: a personal AI project assistant, a portfolio website with a hidden Asteroids game, and a full-stack fictional restaurant brand with an interactive ordering demo. I want to talk about how that happened, what it cost, and what it says about the current moment in software development.
What Got Built
A Python-based RAG pipeline that reads my Obsidian vault, retrieves relevant notes via keyword scoring, and calls the Anthropic API to answer questions about my projects and generate documents. It writes back to the vault with a confirm-before-write review step. Served via a Flask web UI that launches from a batch file. Built from scratch in an afternoon.
A plain HTML/CSS/JS portfolio site with a dynamic content feed pulling from YouTube RSS and a local JSON file, an interactive work timeline with chronological toggle, a contact form, floating illustrated rocks and forest margin trees, and a hidden Asteroids-style easter egg game triggered by clicking the profile image. Deployed to Vercel with a custom GoDaddy domain.
A fictional breakfast restaurant with a full brand identity, marketing website, interactive menu with customization modals and a persistent cart, a checkout flow with payment animation, and a scripted AI ordering demo that walks a complex multi-item order through parsing, confirmation, payment, receipt, and pickup address — end to end. No live API, no backend, zero ongoing cost.
Time
I want to be honest about time because I think people misunderstand what "building with AI" actually looks like. This wasn't prompt-and-done. Each project involved real decisions, real debugging, and real iteration. What AI changes is the ratio of thinking to typing — and it shifts it dramatically in favor of thinking.
My active time across both days was probably 2-3 hours total: designing in Figma, generating and preparing assets, making judgment calls on direction, testing things in the browser, and handling the environment setup that only I could do. The rest was conversation — describing what I wanted, reviewing what came back, pushing back when it wasn't right, and approving when it was.
The AI handled the volume work. I handled the decisions. That's the division of labor that makes this kind of output possible in two days instead of two months.
Cost
Here's the full bill:
| Item | Cost | Notes |
|---|---|---|
| Claude subscription (Pro) | ~$20/mo | Used on Low — Sonnet 4.6. The model doing the work. |
| Anthropic API credits (Hammy) | <$20 | Haiku 4.5 at $0.01/query. Hundreds of queries, pennies each. |
| Figma subscription | ~$15/mo | Design and asset generation for Atti Cakes. |
| GoDaddy domain (attisomething.com) | ~$12/yr | Custom domain. One-time annual cost. |
| Vercel hosting | $0 | Free tier covers everything. Automatic deploys from GitHub. |
| Total (first month) | ~$67 | Monthly after that: ~$35 plus pennies in API costs. |
For less than the cost of a single hour of freelance development, I have a live AI tool, a deployed portfolio site, and a full product demo. The marginal cost of each additional project is close to zero.
The old constraint was that building was expensive. You needed developers, or time, or both. That constraint is dissolving. What's left is the constraint that was always more important anyway: knowing what to build and why.
Quality
Speed without quality isn't useful. So let me be specific about what made the quality hold up across 25 commits and two days of iteration.
The single biggest factor was QA discipline. Every time something shipped I tested it immediately — clicked through the flow, checked edge cases, noted what broke. When something was wrong I could describe it precisely: not "the cart doesn't work" but "the quantity cap on the modal doesn't account for items already in the cart from a previous add." That specificity is what gets bugs fixed on the first try instead of the fourth.
The second factor was building local and deploying often. Every change was visible in a browser within seconds of being written. There was no build step, no staging environment, no ticket queue. The feedback loop was: change, save, refresh, see it. When you can iterate that fast, quality compounds rather than degrading.
The third factor was not trying to do everything at once. Each project had a clear scope. Hammy is a CLI tool that reads markdown and calls an API. attiSomething is a static site with a YouTube feed. Atti Cakes is a multi-page interactive demo. None of them are trying to be more than they are, which means none of them have the debt that comes from scope creep.
What This Actually Means
I'm not writing this to say AI is magic or that experience doesn't matter. The opposite is true. The reason this worked in two days is that I came into it with real knowledge of project management, QA, product thinking, and what good software looks like. AI amplified that. It didn't replace it.
What I think is genuinely new is the accessibility of the amplification. You don't need a team anymore to ship something that looks and works like a team built it. You need taste, judgment, and the discipline to keep iterating until it's right.
25 commits. Two days. Three projects. One subscription and a $20 API budget.
Building is cheap now. The question is what you decide to build.