Question 1

Why can non-technical founders now build software with AI?

Accepted Answer

Coding models crossed a quality threshold in late 2024, and a wave of tools (Lovable, Replit, Bolt, Emergent, Google AI Studio) now turn plain-English descriptions into working apps, with hosting, a database and sign-in handled for you. Andrej Karpathy named the shift 'vibe coding' in February 2025, and it became Collins Dictionary's Word of the Year within nine months. The practical upshot: describing what you want is now enough to get a real, testable product, which is why Lovable says 80% of the people building on it are in non-technical roles.

Question 2

What tools should I start with (Cursor, Claude Code, Lovable, Replit, Bolt, v0)?

Accepted Answer

Match the tool to your comfort level: Lovable, Replit, Bolt and Emergent build and host full apps from plain English in the browser, and v0 generates polished UI. Claude Code, Codex and Cursor are coding agents that assume you will look at code; Claude Code comes with a Claude plan and Codex with a ChatGPT plan, and Codex keeps working in the cloud after you close the laptop. Most non-technical founders start on Lovable or Replit, then graduate to a coding agent. Switching later is easy because your code exports to GitHub.

Question 3

How do real founders ship an MVP with AI in days instead of months?

Accepted Answer

The pattern is consistent: scope one core workflow, prompt a working prototype in hours, put it in front of 5-10 real users within days, and iterate on live feedback instead of specs. Pieter Levels built a flight simulator in 3 hours that hit $1M ARR in 17 days, and Karpathy shipped a paid app (MenuGen) without reading the code. The winners treat the MVP as a test of demand, not a finished product, and expect the messy 30% (auth, payments, polish) to take longer than the demo.

Question 4

What are the limits, when do I still need a real engineer?

Accepted Answer

Vibe coding gets you roughly 70% of the way: prototypes, internal tools, and early MVPs. The last 30% (security, scale, complex integrations, gnarly bugs) is where AI-generated code falls apart, and where thousands of startups are now paying $50K-500K for rescue engineering. Bring in an engineer before real user data, real money, or real traffic arrives, and treat every AI output touching production as unreviewed until a human has read it. Newer tools build the split in: Replit's Cloud Agents test a change and an engineer approves it.

Question 5

What actually counts as vibe coding, and how is it different from serious AI-assisted engineering?

Accepted Answer

Karpathy's original meaning was specific: let the AI write everything and do not even read the code. Simon Willison drew the now-standard line: if you review, test, and can explain the code, that is just software development with AI assistance, not vibe coding. Most founders should treat it as a spectrum, pure vibes for throwaway prototypes and progressively more review as real users and money show up.

Question 6

How do I plan and prompt so the AI builds what I actually want?

Accepted Answer

The single biggest upgrade is to stop prompting features and start with a spec: brainstorm the idea into a written plan (a mini PRD), break it into small steps, then feed the AI one step at a time. Harper Reed's spec-then-plan-then-execute loop and Lovable's Prompting Bible are the two canonical playbooks. Plan in your tool's chat or plan mode, free on Lovable since September 2026. Small, specific, one-thing-at-a-time prompts with real context beat long wishlists, and they burn far fewer credits.

Question 7

How do I keep my vibe-coded app from getting hacked?

Accepted Answer

AI tools routinely ship apps with API keys in the frontend and database security (like Supabase Row Level Security) switched off; scans found roughly 1 in 10 vibe-coded apps leaking credentials. Builders now scan for much of this: Lovable checks every app before it publishes, and Replit can probe yours the way an attacker would. Run those scans and fix what they flag, keep secrets out of client code, turn on row-level rules for every table, and keep projects private. Real casualties (Enrichlead, Moltbook) show attackers actively hunt vibe-coded apps.

Question 8

How do I get my vibe-coded prototype ready for real users?

Accepted Answer

A demo that works for you and a product that survives strangers are different animals: before launch you need real authentication, database rules, environment variables for secrets, error handling, and a deploy setup that separates test from production. Even Karpathy found that deploying his vibe-coded app was 'a painful slog' compared to the exhilarating prototype. Work through a production checklist, run the built-in security scans, and hand-walk every signup and payment path yourself before inviting users. Once live, turn on your platform's monitoring so you hear about failures before users do.

Question 9

How do I add accounts, a database, and payments to my vibe-coded app?

Accepted Answer

Most builders now come with the backend built in: Lovable's built-in backend adds a database, sign-in and file storage as soon as you ask for them, and Replit and Base44 do the same. Supabase is the usual choice when you want a backend you connect and own, and Stripe is how almost everyone takes payments, wired in from a prompt. You do not need to understand Postgres; you do need to know the concepts, one table per kind of thing, row-level rules for who sees what, and secret keys that never live in the frontend.

Question 10

What do I do when the AI gets stuck, loops, or breaks things it already built?

Accepted Answer

The 'fix this, break that' loop is the number one vibe-coding frustration, and it is usually a context problem: the session's memory is full and the AI is patching symptoms. The escape moves are mechanical: stop, start a completely fresh chat, describe the bug and what you have tried, ask for the root cause rather than a fix, and paste actual error messages from the browser console or logs. Combine that with a checkpoint or Git commit before every AI change so a bad fix costs you minutes, not days.

Question 11

How much does vibe coding really cost, and how do I stop burning credits?

Accepted Answer

Plan on $20-100 a month while building seriously. App builders charge by usage: Lovable in credits from $25 a month, with chat free for planning, Bolt in tokens, and Replit from $20 with a Free Mode for everyday tasks. Claude Code comes with Claude Pro at $20 and Codex with ChatGPT Plus at $20; Cursor's cheaper plans now cover only Grok models, with other makers' models billed at API price. The hidden cost is sloppy prompting: loose iteration can cost 3-4x the credits of disciplined prompting. Start cheap, plan before you prompt, and upgrade only when you hit limits.

Question 12

Do I still need to learn to code, or is prompting enough?

Accepted Answer

You no longer need syntax, but you very much need computational thinking: how to break problems down, what a database or API is, and how to read what the AI did well enough to catch nonsense. Replit's CEO says stop learning to code and start learning to think clearly with machines; Andrew Ng counters that founders who understand code direct AI far better. The pragmatic answer: build first, and let each thing that breaks teach you the concept behind it.

Question 13

Can vibe-coded products become real businesses?

Accepted Answer

Yes, and the evidence is no longer anecdotal: Base44 sold to Wix for $80M six months after a solo founder started it, Pieter Levels' vibe-coded flight sim hit $1M ARR in 17 days, and India's Emergent went from zero to $100M ARR in eight months. The tools themselves are among the fastest-growing software companies ever (Lovable raised $400M at a $13.3B valuation in August 2026), which tells you where demand sits. The catch: code is now cheap, so distribution, retention, and a real problem worth paying for are the actual moats, which is why Lovable now helps you market what you build.

Question 14

Can I vibe code a mobile app and get it into the App Store?

Accepted Answer

Yes: Replit's agent builds native mobile apps you can test on your phone and submit for App Store review, Google AI Studio builds native Android apps you can publish to the Play Store for testing, and Emergent, Vibecode and Rork generate mobile apps from a prompt. Mobile is harder than web, though; Apple's review process, device testing, and native quirks add friction no AI removes yet. Most founders validate as a web app first and go native once users prove they want it in their pocket.

Question 15

How do I avoid losing work? Version control for founders who don't code

Accepted Answer

The most common regret among new builders is losing a working version they cannot get back. The insurance policy is Git: think of it as unlimited save points; commit (or checkpoint) before every AI change so a bad session costs minutes, not days. You do not need the command line: connect your builder's GitHub sync or use GitHub Desktop, try risky changes in a copy first (Lovable calls these drafts), and make 'commit when it works' as automatic as saving a document.

How founders use AI for Vibe Coding

The questions

Also in Starting Up