Lovable

Lovable.dev — Platform Skill

Lovable is a launchpad, not a development platform. It generates React full-stack apps from natural language prompts — fast, clean enough for prototypes, not built for scaling or ongoing development.

The single most important thing to understand: Lovable is Step 1 only. Never Step 2+. Once you export to GitHub and start developing elsewhere, you don't bring it back.

The Gate That Governs Everything

Lovable CAN export TO GitHub ✅ Lovable CANNOT import FROM an existing GitHub repo ❌

This is the hard constraint that determines Lovable's position in any pipeline.

Workaround: Zip from another tool → import zip into Lovable. One-time operation, not a repeatable workflow.

What this means:

• Lovable is always the starting point for net-new projects
• It cannot participate in iteration on existing codebases
• The universal pattern is: Lovable → GitHub → real dev tools (Cursor, Claude Code)
• Planning to use Lovable? Make sure nothing exists yet that needs to be imported

What It Generates

From a natural language prompt, Lovable produces:

• React frontend (routing, layouts, components, basic UI)
• Backend API routes
• Database schema (via Supabase or Lovable Cloud)
• Authentication flows
• Basic CRUD operations

The output is a working full-stack web app. Not production-grade, but clickable and demonstrable. Good enough to validate an idea with real users.

Built-in Tools

• Visual Editor: Click elements in the preview to adjust color/size/spacing without prompts
• Chat Mode: Discuss problems or explain code without consuming credits
• Code Mode (paid only): Edit generated code directly within platform
• Figma Import: Convert Figma designs into working components
• One-click Deploy: Publish to Lovable subdomain or custom domain (paid)
• GitHub Sync: Two-way sync to a GitHub repo — you always own the code

Backend Options

Two paths for backend functionality:

1. Supabase integration: Auth, database (Postgres), storage. Requires separate Supabase account. More mature, more control, but you manage two platforms
2. Lovable Cloud: Built-in backend (database, storage, auth, pre-integrated AI models). Newer, simpler, less customizable. Good for "just make it work" prototypes

Other integrations: Stripe, Clerk, OpenAI, Netlify deployment.

Credit System & Pricing

Lovable runs on credits. Every AI interaction costs credits. This is where it gets expensive if you're not careful.

Credit consumption rules:

• Every prompt = credits consumed (amount varies by complexity)
• Simple edits: fewer credits. New pages/refactoring: more credits
• "Fix" messages do NOT count toward quota — use this aggressively for debugging
• Monthly credits roll over. Daily credits do not
• Debugging loops are the #1 credit drain — the AI tries to fix a bug, fails, reintroduces old errors, each attempt burning credits

Cost reality: The sticker price is misleading. Complex projects easily burn through Starter credits in a few sessions. Budget for Launch or Scale if doing anything real.

For full pricing details, read references/pricing-and-limits.md.

Where Lovable Works

Use Lovable when ALL of these are true:

• The project is brand new (nothing to import)
• You need a working prototype, not a production system
• The complexity is low-to-moderate (CRUD, dashboards, landing pages)
• You accept that ongoing development happens elsewhere after export

Good use cases:

• MVP for idea validation
• Clickable prototype for stakeholder demos
• UI scaffolding to export and refine in Cursor/Claude Code
• Landing pages, marketing sites, email capture pages
• Simple internal tools where the lifecycle is: build → deploy → done
• Hackathon projects

Where Lovable Fails

Architecture problems:

• Generated code is not built for scaling. Data structures are inflexible
• Logic gets tightly coupled in ways that break when requirements change
• Small feature changes can require rewriting large generated sections

Complexity ceiling:

• Complex backend logic (user roles, multi-tenant auth, complex business rules) → breaks
• Enterprise apps with payments + distributed systems + multiple API integrations → no
• Real-time messaging, recommendation engines, custom data pipelines → no
• Anything requiring custom backend architecture → no

The debugging loop: The most common complaint. The AI gets stuck trying to fix bugs — tries a fix, fails, re-introduces an old error, burns credits. At some point you must export and fix manually. This is not a bug, it's a fundamental limitation of the approach.

Code quality:

• AI-generated code can be hard to debug manually
• Not optimized for performance
• Generic UI patterns — limited visual distinctiveness
• May not follow framework best practices

Pipeline Doctrine

Lovable (generate MVP) → GitHub (export) → Cursor / Claude Code (real development)

Lovable occupies exactly one position in the pipeline: the first step for net-new projects that don't yet exist. It is never a middle step. It is never used on existing repos. The moment you need ongoing development, maintenance, complex features, or production reliability — you're in Cursor/Claude Code territory.

Fleet Reference — Ultimate Tool (UT)

The Ultimate Tool (templates/snap-on/ultimate-tool/) intersects with Lovable at the pipeline boundary. Lovable generates the initial UI for child repo frontends, which are then exported to GitHub and developed further with Claude Code.

How UT and Lovable relate:

• Lovable builds the frontend scaffolding for child repos that need a UI
• Those frontends call UT via HTTP through the Cloudflare Worker rim (SH-06)
• Lovable itself is never part of UT's runtime — it's Step 1 only
• After export, the frontend code lives in the child repo and is maintained by Claude Code

Pipeline position:

Lovable (generate UI) → GitHub (export) → Claude Code (develop) → UT (call via HTTP)

The UT dashboard (fleet control panel) in dashboard/ was NOT built with Lovable — it was built directly with Claude Code. Lovable is reserved for net-new child repo UIs where rapid prototyping justifies the credit cost.

UT's full spec: templates/snap-on/ultimate-tool/README.md

Dave's Operational Notes

Known Failure Modes