Healthy path · Step 5 Later · when real users appear

Healthy path · Step 5 (later — not night one)

When real people use it

What this step is

Steps 1–4 = learn, pick a tool, ship a sandbox win.
Step 5 = the gate before strangers, private data, or money rely on what you built.

Reading this early is fine. Treating a night-one demo as “production” is not. Pretty is not the same as safe.

No sandbox win yet? Go back: 1 · 2 · 3 · 4.

The 80% wall

AI is strong on the first stretch: something that looks finished. The hard stretch is auth, private data, payments, edge cases, and keeping it running. That gap is the 80% wall — judgment and checks, not one more mega-prompt.

Signs you’ve hit the wall optional
  • Login “works” until two users can see each other’s data
  • Keys end up in chat, screenshots, or client-side config
  • Payments stuck in test mode forever
  • Layout looks fine; network errors hang or leak details
  • Dependencies appeared “because the agent added them”

What usually breaks first

Pattern → why vibe-built code hits it → first move. Terms in brackets are industry names for later reading.

What goes wrong Why AI often does this First move
User A can open user B’s records (access control / IDOR) Models invent “simple” queries and forget ownership checks Every read and write must prove the signed-in user owns that row
Secrets leak (API keys, tokens) Keys pasted into chat or shipped in front-end code Revoke/replace the key; keep secrets server-side only; never commit them
Unsafe HTML or SQL (XSS / injection) String-built queries and raw HTML from AI “just works” demos Use the platform’s safe query APIs; don’t trust unescaped user HTML
Dependency sprawl The agent adds packages “to help” Require a reason for each package; lock versions; delete unused ones
No recovery path Everything lives only inside a builder sandbox Export source; know how to rebuild and redeploy without that one laptop

Common checks before real users

Informal learning checklist — not a security audit, certification, or guarantee. Use it to ask better questions; get professional help for real personal data, payments, or regulated work. Printable companion: First production checklist.

Five high-impact topics to understand

More gate items when you go further
  • Database rules — default deny; test as two users (e.g. Supabase RLS)
  • Injection / XSS — no string-concat SQL; careful with AI-generated HTML
  • Uploads — type/size limits; no executables as public files
  • Dependencies — pin versions; remove unused agent packages
  • Admin — not “a secret URL nobody will guess”
  • Logging — no passwords, tokens, or full card numbers
  • Backups / monitor — you know what happens if it dies on a Friday night

Quick checks before you advertise

  1. One happy path for the money or data change that actually matters
  2. One negative check: user A cannot read user B’s record
  3. Empty forms + narrow (phone) window by hand
  4. A staging URL with real-ish settings — not only “works on my laptop”

Leaving the builder sandbox

When a hosted builder is no longer enough for real users:

  1. Source in Git with a README and example env (no real secrets in the file)
  2. You (or CI) can rebuild without one magic machine
  3. Dev / staging / prod settings are separate
  4. Monitoring and backups before hard marketing push
  5. AI workflow: small, reviewed changes — not “rewrite the whole app”

When to stop accepting AI rewrites

Pause big AI diffs when you cannot explain them, when security-sensitive files change “by accident,” or when tests get deleted to force a green build. Undo, restate the rules, re-apply in thin slices. Longer note: When to stop accepting AI diffs.

Healthy path · later gate

You know where the gate is

Night one was Steps 1–4. This page is the later gate — educational notes only, not a certificate that an app is safe for customers.

If you already finished the lab path, you’re done with the beginner map. Come back here when a demo is about to hold real people or real money — and get professional help for serious risk. Disclaimer · Terms

Site by Ian Piggott. If this path helped, tell one person stuck on “build me an app.” About · Contact

Deeper when you’re actually shipping

Not night-one homework. Educational only — see the notice at the top of this page.