Why we built the Vibe Coder's Guide to MVP
A friend of mine spent six weekends building "her thing." A meal planner. Nothing crazy — pick what you want to eat for the week, get a grocery list. She used Claude Code. She had a working prototype by the end of weekend two.
She never shipped it.
When I asked why, she didn't say the code was broken (it wasn't), or that her idea was bad (it isn't), or that she ran out of time. She said this: "I don't know if it's actually good enough."
She had no auth, no privacy policy, no domain, no analytics, no error pages, no idea what to do about cookies, no clue whether her AI feature would survive the OpenAI bill, and no framework for asking herself "what does it look like to be done?" So she didn't ship. The code sits in a folder.
This is the most common shape of failure for vibe coders right now, and I think the gap is bigger than most people realize.
The agents are good enough. The confidence isn't.
I've spent the last year watching Claude Code and Codex go from "interesting demos" to "actually competent senior developers in your terminal." When directed well, the code these agents write is shippable. They handle TypeScript, they understand modern frameworks, they write tests, they refactor with judgment. The artifact is real.
What's missing is not on the agent side. It's the rubric.
A senior engineer doesn't ship a prototype. They ship something that has been quietly hardened against the dozen things that could embarrass the founder later: WCAG accessibility violations, missing security headers, a license file in the repo, a real privacy policy that mentions Stripe and OpenAI by name, an admin password instead of "admin/admin", a 404 page that isn't a Next.js stack trace. None of these are hard. All of them get skipped because nobody told the agent to do them.
That's it. That's the gap.
What we built
The Vibe Coder's Guide to MVP is a stack of seventeen markdown files. Each one is a "skill" — a structured set of instructions an AI coding agent reads and follows. The agent has a conversation with you about the parts that need your judgment ("which auth provider?" "what should the privacy email be?") and does the rest itself ("install the headers, write the consent flow, ship the file").
Together they walk a project from "I have an idea" to "it's deployed, it's secure, it's accessible, the legal pages exist, and you have a one-pager I can hand to an investor."
You can read them all on this site. You can also download the bundle and drop it into Claude Code or Codex yourself — git clone https://github.com/titan-alpha/vibe-coders-guide-to-mvp-skills and your agent inhales the rubric in one shot.
Three principles I keep coming back to
One. Plain language wins. The audience is not engineers. The skills don't say "denormalize the schema for read-heavy access patterns." They say "duplicate a couple of fields between tables so reads are faster." If a sentence requires you to already know the answer, it's not teaching anything.
Two. Always default to the simplest viable solution. If polling at 30 seconds works, don't reach for WebSockets. If a footer link covers Terms of Service, don't build a clickwrap. If a single $0.50/month domain on Route 53 works, don't propose Kubernetes. Most "engineering rigor" at the MVP stage is performative — a sophisticated way to not ship.
Three. Pause before destructive or paid actions. The agent always asks before it deletes files, drops tables, deploys to prod, or spends money. This sounds obvious. It isn't, and the skills are explicit about it because the asymmetry matters: a 10-second confirmation prevents a 4-hour recovery.
Where this is going
The skills are version one. They will get better as people use them and find the parts that are wrong. They will get worse if I add them as I think of them — half the discipline is keeping them tight.
I'd like, eventually, to see this become the thing that someone with an idea downloads on a Saturday morning, finishes by Sunday night, and has shipped on Monday. Not because the agent did it for them — because the agent helped them do it.
That's the MVP of the vibe coder, by the vibe coder, for the vibe coder.
If you're sitting on a folder of code that "almost works," I want you to ship it. The bar is lower than you think. The hardest part is convincing yourself you're allowed to.
Believe you can. You can.