Building Web Apps with AI
Vibe-coded apps from the community: tools, patterns, and pitfalls.
On this page
Beyond simple flyers and letters: actual web applications. The chat's maturation arc runs from "holy cow, you can build apps in minutes" to "here's what to vibe-code and what not to."
The tool decision
| Need | Tool |
|---|---|
| 15-minute prototype, single page, no DB | Lovable |
| Database-backed, multi-screen | Base44 |
| Real engineering, GitHub | Claude Code |
| Already in an IDE | GitHub Copilot ($10/mo) |
| Power-user big codebase | Cursor ($300-500/mo at heavy use) |
| Free phone-based prompting | Google Jules |
Full comparison in Coding Builders.
What to build (safe patterns)
- In-browser, no server, the page reads input, computes, and displays. No database.
- No logins, no accounts.
- No payments, no banking info.
- No customer PII storage.
- Curated corpus, AI re-presents. (See Maamorim App.)
- MCP into vetted backends when you must integrate (Salesforce, Tally, Stripe).
These all live up to the community vibe-coding rules (Vibe Coding).
What not to build (alone, with AI alone)
- Anything user-facing with logins.
- Anything with payment / banking integration without a security review.
- Mission-critical email integrations.
- Mass-messaging bots on personal WhatsApp numbers; see WhatsApp Automation.
- Mass donor/contact databases that AI agents can directly query without auth boundaries.
Real examples (shipped)
See Chabad Apps for the running list. The pattern:
- megillah.app, maamorim.app, mishna.me, mikdash.live
- the berel.me portfolio (a dozen-plus tools)
- learningtanach.org, Tzvi-to-Tzadik, Book of Esther, Rashi Roots Map
- cypcampaign, sms-invite-joy, chabadvocate
- Tenpr.app
- ChabadUp.com (a full app, 100% AI)
- Rebbe's Global Footprint (built in 15 minutes)
- ai770.com/megillah, ShliachFlow
Lessons from the field
- Restart the chat / refactor often. A learningtanach.org post-mortem: "Each addition made the codebase more fragile. By the end, I had six different timing systems contradicting each other."
- Test in real environments early. Especially for ChabadOne (XHTML 1.0); see ChabadOne Integration.
- Use AI to write scripts, not to process data. "In general, I always ask AI to make a script instead of asking it to directly process the data."
- Have AI assess your security. "If you build using Claude Code you simply ask it how it can improve the security…"
- Watch token usage. Claude Sonnet 4.6 ran out mid-project for one member; restarting in a different model caused issues.
Open questions
- How to graduate from a Lovable prototype to maintained software.
- A template that 200 shluchim can install without each becoming a security incident.
- Multi-maintainer vibe-coded projects; most are bus-factor 1.