Best prompts for Lovable, and what Lovable already does

Lovable design systems discover the tokens you already have and enforce them. They police, they do not author. What to paste in, and where to keep it.

A large empty frosted glass frame with a smaller richly detailed color- block panel hovering just in front of it about to slot in, lit coral and teal
An empty frame and the thing that goes in it. Only one of those is supplied.

Lovable's own design guidance is free on every plan and already settles a font pair and a palette. When a prompt leaves the look open, it asks up to three design questions, on color palette, typography and layout, each with up to four options shown as swatches, font samples and wireframes, and the questions cost no extra credits. Alongside it sit a prompt library of 57 example prompts, a twenty style reference, and roughly 15,000 words of prompting and design documentation.

Anyone selling you prompts for Lovable who does not tell you that is selling badly. Start there, for nothing, and come back when you hit the edge of it.

The edge is specific and it is worth knowing before you spend anything:

  • Lovable's design layer names fonts. Yes, as heading and body pairs.
  • It applies exact palette colors. Yes.
  • It names an easing curve. No.
  • It names a duration in milliseconds. No.
  • It handles prefers-reduced-motion. No.

Those last three are what a full brief adds. Each of these designs is one such brief, live, and its prompt pastes into Lovable as it is:

Design systems police your tokens, they do not choose them#

This is the part that gets misread everywhere, and it changes what you should buy. Lovable's documentation is plain about where the tokens come from:

Tokens are discovered from top-level styling files: CSS custom properties in files like src/index.css or src/styles/tokens.css, a tailwind.config, or a theme/tokens source file.

Discovered. Not authored. The design system reads what your project already decided and turns it into a schema.

What it then does with that schema is genuinely deep, and it is the strongest thing in Lovable's design layer:

every generation is scanned for violations of the design system

Four kinds, per the docs: raw color literals where a token should be used, one-off values that bypass the scale, inline styles that override the defaults, and local components for things that should come from the system. When it finds one it retries and corrects before finishing.

That is a per-generation linter for design drift, and no pasted prompt can do it, because a pasted prompt does not run inside the build. It is a real advantage and it is worth paying for on its own terms. It also only enforces what you wrote down, so writing the tokens first, in the W3C format, is the step that gives it something to enforce.

The limits are equally plain, all from the same documentation:

  • One design system per project.
  • React only. Vue, Angular and Svelte are not supported.
  • An existing project cannot be converted into one.
  • system.md is the only free-text art direction slot, hand-authored and preserved across releases.

So the system is infinitely deep and completely empty. It will carry any level of specification you give it, including a very long one, and it supplies none of that specification itself. Arrive with no tokens and no opinion, and the same model that writes your code also invents your palette, which lands you back at the default look the design system was meant to protect you from.

A strict glass grid snapping a set of translucent tiles into perfect alignment, every tile itself blank and colorless, lit coral and teal on the structure only
A fixture holds a part to a tolerance. With no part in it, the tolerance sits idle.

The five worth having open, ranked#

Ranked by what each one supplies that Lovable does not, which is a narrower test than ranking the prompt libraries against each other. The free ones come first, because at your current step they are better value.

SourceWhat it isCost
Lovable's prompt library57 example prompts and a 20 style table, no loginfree
Lovable's own off switchOne documented sentence that skips the design questionsfree
Meez's Lovable hubFour complete prompts inlined, 62 free across the sitefree tier
A skill file you write once100,000 characters, available to every project in the workspacefree
ThanorFull art direction per design, written as valuessome designs free; pricing
  1. Lovable's prompt library. Free, no login, 57 example prompts and a twenty style table of keywords. It names one typeface, Inter, and no hex at all, which is the honest ceiling of free first-party guidance.
  2. The off switch, quoted from Lovable's docs: "To build without design questions, describe the look in your prompt." The docs add that Lovable builds directly when a prompt gives named fonts, colors, a brand or a style. That is the entire use case for this category, published by the vendor.
  3. Meez's free Lovable hub. Four complete prompt texts inlined on the page itself, and 62 prompts free across the site with a free account. Good, long, and free to read this afternoon.
  4. A skill file you write once. Up to 100,000 characters, with no skill-specific credit cost, available to every project in the workspace and applied automatically when a request matches its description. Better delivery than a copy button, and Lovable built it without meaning to compete in this category.
  5. A full art direction brief. Ten blocks, exact values, a named curve with a real duration, a never list, acceptance checks. That is what Thanor writes, one per design, and it is the layer nothing above supplies.

What a prompt has to carry that Lovable never asks for#

Left to its defaults, the Lovable output we sampled in September ran Tailwind's stock easing curve at around 150ms on almost every transition, reached for Inter, and rarely shipped a prefers-reduced-motion rule at all.

None of that is a weakness in the model. Lovable will build whatever motion a detailed prompt asks for, at any curve and any duration you name. These are framework defaults, which is a house style choice, and the reason they show up everywhere is that nobody overrode them.

So a brief that is worth pasting names, at minimum:

  1. Four to six hex values, each with a role: ground, surface, ink, accent, and the hue bias of the neutral stated rather than left to chance.
  2. Two typefaces with the fallback stack written out, and a scale in px.
  3. A named easing curve with its real value and a duration in milliseconds, the part a GSAP ScrollTrigger prompt has to get right.
  4. What is driven continuously by scroll progress against what triggers once.
  5. One signature moment, described as a mechanism rather than as a mood.
  6. A prefers-reduced-motion rule under which the page still makes its point.

The house rule underneath all six: exact values, never adjectives. #0F1317, not "deep charcoal".

Where to put a long brief so Lovable keeps it#

ContainerCeilingAppliedCost
A skill file100,000 charactersautomatically, on description matchno extra credit cost
Project instructions, formerly knowledge10,000 charactersalways included in contextfree
AGENTS.md at repo rootno stated ceilingread with every messagefree
system.md in a released design systemno stated limitevery generation, connected projectspaid plans

Be precise about which one you reach for. A full brief will not fit in project instructions, which cap at 10,000 characters. It fits comfortably in a skill file, and that is where it should live if you want it applied to everything you build in that workspace from now on.

Where this leaves you#

Take Lovable's free docs first. Write your tokens before you connect a design system, so the linter has something real to enforce. Then, if the part you cannot write is the direction itself, that is the gap Thanor exists to fill: the palette, the scale, the curve, the duration and the order, decided per design and written as values you can paste straight into a skill file.

Free accounts open the designs marked free, and the plans open the rest. Open Vesperion, copy its prompt into a Lovable project and compare it with what you would have written; it is free with an account. If the look you are fighting is the vibe-coded one, the same argument runs through vibe design.

Questions this raises

Do Lovable design systems create a design for you?

No. They discover the tokens already present in your project's styling files, render them into a machine readable schema, and then check every generation against them. That is enforcement rather than art direction. Start a project with no tokens and there is nothing for the design system to enforce.

What is the best prompt to use with Lovable?

The one that arrives with decisions already in it: four to six hex values with a role each, two named typefaces with the fallback stack written out, a type scale in px, a named easing curve with a duration in milliseconds, and the section order. Lovable will build whatever you specify. It just will not choose for you.

Does Lovable have a free prompt library?

Yes. Lovable's public documentation includes a prompt library of 57 example prompts, readable with no login, alongside a 20 style reference and roughly 15,000 words of prompting and design documentation. It works in style keywords rather than values, and it is a good, free starting point.

Where do I put a long design brief in Lovable?

In a skill file, which takes up to 100,000 characters and carries no extra credit cost of its own, or in system.md inside a released design system on a paid plan. Project instructions, which Lovable used to call knowledge, are capped at 10,000 characters, which is not enough for a full brief.

Part of

What is on the market

The whole field in one place, counted rather than ranked on a hunch. Who sells what, at what price, and which of them publishes a number you can check.