One-page playbook. Lock scope to one screen, one user, no auth, localStorage-or-bundled-cloud [1] [2]. Default tool is Lovable with Bolt.new as the warm backup [3] [4]. Spend 30 min refining a PRD in ChatGPT (free, no credits burned), then paste once into Lovable Agent mode [5] [6]. Sign-ups happen 24-48 h before the room, off the venue wifi, with a phone hotspot in reserve [7] [8]. Take-home failures are not the deploy click — they are the Supabase 7-day pause [9] and OAuth refresh-token expiry on day 8 [10]; the printed take-home card should pre-empt both. Host budget for 20 attendees: ~€0 free-path, ~€23 (≈$25) comfort-path with one shared Lovable Pro seat [11].
The cross-cutting picks
Across the four research angles, four decisions repeat — they are the spine of the workshop:
-
Scope is the lever, not the tool. The app-ideas survey and tool-stack survey independently land on the same constraint: marketplaces, OAuth login, payments, native iOS, and real-time multiplayer all blow the 2-3 h budget regardless of which builder you pick [12] [13]. Cut to a single-user, single-screen rectangle and the builder choice almost stops mattering [14].
-
Lovable + Bolt is the redundant pair. Lovable wins on “looks done” and SSR-light bandwidth [15]; Bolt wins on signup friction and walk-in attendees [16]. They fail in opposite ways too — Lovable has logged 314+ outages since May 2025 [17], Bolt’s WebContainers break on ad-blockers, VPNs, and non-Chromium browsers [18]. Keep both warm.
-
Plan in ChatGPT, build in Lovable — physical separation enforces the rule. Lovable’s free tier is 5 daily / 30 monthly credits [19] — a 20-person cohort has ~600/month to share, ~2/attendee for a landing page and ~0.5 per surgical edit [20]. Refining in ChatGPT first is free and pre-empts the “vague prompt → vague app → debug-loop credit burn” cycle [5]. Drill Lovable’s edit template — “Change [X] to [Y]. Keep [everything else].” — and the three-strike rule (revert after two failed corrections) [6] [21].
-
The take-home card is the most under-built artifact. Every child page touches this and the conclusion converges: the URL exists at minute 145, but Supabase pauses after 7 days of DB inactivity [9], Google OAuth refresh tokens die after 7 days in Testing mode [10], and Lovable apps regularly leak API keys baked into the JS bundle [22]. Bundled clouds (Bolt Cloud, Lovable Cloud) sidestep the SQLite-on-Vercel data-loss trap [23] [24] — but the printed card still needs the “if it says paused, click resume; data is safe” sentence [25].
Where the angles pull against each other
The deploy angle and the tool angle disagree on the default. The tool stack research puts Lovable first on workshop-day UX (chat-only, polished output, classroom track) [26]; the take-home deployment research puts Bolt.new + Bolt Cloud first on the two-month horizon (1M tokens/month, no sleep, bundled DB, no card) [27] [24]. Resolution: use Lovable in-room for the planning-to-build UX win, but pick the bundled-cloud path inside Lovable (Lovable Cloud auto-pauses on $0 rather than going broken [28]) — and reserve Bolt + Bolt Cloud for the participants who explicitly want a longer-lived URL on a more generous free tier.
The open questions
No first-person, dollar-itemised post-mortem of a vibe-coding workshop killed by conference wifi was found — wifi guidance is industry-vendor framing, not a primary write-up [8]. No published template for the take-home card exists either; the recommended contents are synthesised from failure-mode posts [29] [30]. If you run this playbook, the highest-value thing you can publish is a post-mortem with spend in EUR, wifi survival rate, and the percentage of URLs still live at day 14 — that is the gap a future Scout run would close.