Vibe Coding Is Real. Most Advice About It Is Not.
Somewhere between "AI will replace all programmers" and "AI code is unusable slop" is where actual work gets done. This series lives there.
Vibe coding — describing what you want in natural language and letting an agent build it — is genuinely transformative. We built and deployed a full underwater ocean game, with IMU motion controls on mobile, in about five minutes (that story is Part 2). That's not a demo trick. It's live at a real URL, and people play it on their phones.
But here's what the hype cycle won't tell you: the prompt is the easy part. What separates a vibe-coded toy from shipped software is everything that happens after the first generation — the refinement, the deployment, the maintenance, the iteration. That's where this guide earns its keep.
The Thesis
Our full workflow, which the rest of this series walks through step by step:
- Bootstrap with Quests — an open source, local-first desktop agent. Zero credit anxiety, any LLM you want, real code on your real filesystem. (Part 2)
- Hand off to VS Code + Claude Code — for fit, finish, architecture cleanup, and the engineering judgment that turns a prototype into a product. (Part 3)
- Deploy it yourself — because platforms are great until they aren't, and if you've deployed apps for decades, you already have superpowers the hosted builders can't sell you. (Part 4)
- Delegate maintenance to your autonomous twin — NanoClaw and PicoClaw running PM triage, bug fixes, feature branches, and testing while you sleep. (Part 5)
- Take it past the browser — the open source payoff: the same stack, adapted for IoT and edge devices. (Part 6)
Each stage uses the tool that's best at that stage. No single tool does the whole journey well, and anyone telling you otherwise is selling you a subscription.
The Landscape, Honestly
The "prompt-to-app" space in 2026 breaks into a few camps. Here's the map, with our honest read on each:
Hosted chat builders (Lovable, Bolt.new)
Chat-first, browser-based, batteries included. You describe an app, they generate a full-stack React/TypeScript project with backend, auth, and one-click deploy.
What they're great at: zero-friction ideation. Non-technical stakeholders can participate. A founder can validate an idea before lunch.
Where they hurt: credits. Heavy iteration — and debugging especially — burns credits fast, and the meter creates a subtle pressure to accept "good enough" output. Repeated broad prompts tend to degrade the codebase over time: inconsistent patterns, cascade breakage when you change a feature, hallucinated abstractions. And they're weakest exactly where experienced builders are strongest: existing large codebases and custom deployment.
Hosted IDEs and agents (Replit Agent, v0, Cursor)
- Replit Agent: full hosted dev environment, broad language support. Great if you're fine living in someone else's cloud.
- v0: excellent design-led React/Next.js generation. Frontend-first by design.
- Cursor: the strongest option for working inside an existing large repo — but it's a proprietary IDE, not a from-scratch builder.
Open source, local-first agents (Quests)
This is our bootstrap layer, and the reason this guide exists.
Quests (quests.dev, Apache 2.0) is a desktop app — Mac, Windows, Linux — with a browser-like tabbed interface wrapped around a serious coding agent. You prompt; it performs targeted file edits with real-time linting and error correction. Multi-project tabs. Built-in versioning. Bring your own LLM: OpenAI, Anthropic, Groq, DeepSeek, Ollama running locally, anything OpenAI-compatible.
Why it wins for experienced builders:
- No credit anxiety. The core is free open source. You pay only your chosen model's API costs — which can approach zero with local models or cheap providers.
- Real files, real Git, real machine. The output is a standard project (Vite + Hono + oRPC + React by default; other stacks supported) sitting on your disk. Nothing between you and the code.
- Model freedom. Point it at the best model for the job today, not whatever the platform's margin allows.
- Open source means adaptable. This is the sleeper feature. Because you can read and modify the agent itself, the same tooling philosophy extends to work no hosted builder will ever touch — IoT, edge devices, ancillary hardware. (Part 6 is entirely about this.)
Honest caveat: Quests is younger than the hosted incumbents. Fewer no-code polish features, and you manage your own infrastructure. If you've been deploying large apps for decades, that "caveat" is actually the pitch.
The One-Table Version
| Lovable / Bolt | Replit / Cursor | Quests | |
|---|---|---|---|
| Where code lives | Their cloud | Their cloud / your repo | Your machine |
| Model choice | Theirs | Mostly theirs | Any (BYOK, local) |
| Cost model | Credits (spirals) | Subscription | Free + your API costs |
| Best phase | Idea validation | Deep editing | Bootstrap → handoff |
| Lock-in risk | High | Medium | ~Zero |
| Existing large codebases | Weak | Cursor: strong | Fine (it's just files) |
The Rule of the Guide
Use the fastest tool that leaves you owning the result.
Speed without ownership is a demo. Ownership without speed is the old world. The rest of this series is the practical path to having both.
Next: Part 2 — Zero to Deployed in Five Minutes, where we build Abyss Drift, an underwater game with mobile motion controls, and put it on the internet before the coffee gets cold.