SaaS MVP with user auth
Why it works: Lovable asked 4 questions (audience, workspace naming, invite delivery, look), then built and tested the invite flow itself as a second user.
Open this prompt page →8 weak prompts rewritten into ones Lovable builds right the first time, 3 of them measured in credits. Then check your own prompt against the same seven rules before you spend credits on it.
By Marco Kohns · Every measured example was run in our own Lovable workspace
Check your prompt
Runs in your browser. Nothing is sent or saved.
Who it is for found
Named components missing
Add: Name parts: table, card, modal, form, chart.
Empty, loading, error states missing
Add: Show an empty state, a loading skeleton and a retryable error.
Visual direction missing
Add: [Heading font], [body font], [2 colours], 8 px spacing.
Guardrails missing
Add: Do not change [what must stay]. Only edit [where].
Questions first missing
Add: Before building, ask me up to 3 questions. Wait for answers.
No vague adjectives missing
Add: Replace modern or sleek with concrete fonts, spacing and components.
Examples
Why it works: Lovable asked 4 questions (audience, workspace naming, invite delivery, look), then built and tested the invite flow itself as a second user.
Open this prompt page →Why it works: All five features came back in one round. Lovable also invented a brand name for the page, which is why the prompt scopes proof to placeholders: name yours.
Open this prompt page →Why it works: Lovable asked 2 questions, built 4 base components and a gallery, then tested slow loading, failure and retry itself.
UI prompts, with the gallery screenshots →Why it works: Naming server-side and URL state stops Lovable from loading every row into the browser, and makes the view shareable and back-button-safe.
The table prompt →Why it works: What to build, where it goes and what must stay untouched: the three parts Lovable's prompting guide asks for in a change prompt.
Common mistakes →Why it works: Concrete values replace adjectives the model cannot act on, and a committed visual direction skips Lovable's design questions. Project knowledge keeps later sections consistent.
UI prompts →Why it works: Expected versus actual, where to look, and a guardrail. A vague fix request costs a build message each time; a Plan mode message costs 1 credit and changes nothing until approved.
Common mistakes →Why it works: Missing row-level security is the flaw behind CVE-2025-48757 in Lovable projects. The docs say Lovable sets up access rules for features that store user data; the prompt makes the rule explicit so the generated policy is the one you meant.
Backend prompts →Measured runs
Pasted unedited into fresh projects in our own workspace, credits read from the balance before and after. One screenshot per run; the full sets are on the prompt and category pages.



How long should a prompt be?
Long enough to name screens, states and guardrails. Short enough to read before you paste. Change prompts are far shorter: one change, where, and what stays.
| Change prompt (the example above) | 23 | |
|---|---|---|
| Shortest build prompt | 298 | |
| Median build prompt | 541 | |
| Longest build prompt | 550 |
Word counts across the 122 base prompts in the library, computed at build time. The change prompt is the Change one thing example above.
Lovable's own advice, applied
Product, audience, why they use it, the key action. Then the look.
Shared states, dialogs and navigation before screens.
One component at a time, named parts, final copy.
One change per message. Keep what works.
Phase names from Lovable docs, Prompting best practices, read 22 October 2026. The example lines are ours.
Every library prompt carries all seven rules. Free to copy, 122 of them.
Who it is for, named components, the empty, loading and error states, a visual direction, guardrails, and a request to ask questions first, with no adjectives the model cannot act on. The checker above tests for those seven, and every library prompt carries them.
Lovable publishes no per-prompt figure; Build mode is billed on the work done. Our measured runs on library prompts cost 1.8 credits for a landing page, 9.5 for a states system and 10.4 for a SaaS foundation with auth, read from our own balance before and after. Lovable's illustrative example for a small style change is 0.50.
Lovable's guide says a full-page prompt gets you noise and a section-based prompt gets you signal. Build the frame and the shared states first, then one screen at a time. The library prompts are one message each but carry a build order, which is what makes one paste behave like a sequence.
The three measured ones are abridged here; copy the full prompt from its page, replace the bracketed parts, and paste. The other five are patterns to adapt. Leave a bracket in and Lovable invents what goes there.
It fills the gaps with your answers instead of guesses and saves them to project knowledge for later prompts. In our runs the closing line produced 0 questions (before the line existed), then 2, then 4, each costing nothing extra.
Docs read 22 October 2026.
Written by

Founder of ProtoBites - Venture Growth Studio
Growth PM at a Silicon Valley scale-up (a16z and General Catalyst backed), ex-Techstars where he consulted 13 early-stage startups, Reforge-trained. Every prompt on this site comes out of shipping ProtoBites' own portfolio products.