NightMarquee alternatives: four variants per prompt

NightMarquee sells 36 prompts, six of them free, and writes each one four times over: once for Lovable, once for Bolt, once for Cursor, once for Claude. Six complete prompts, all four variants each, are readable with no account.

That last part decides most of this page. You can check the product before you form an opinion about it, which is not true of much of this category.

Below is what is in the box, what the previews do and do not tell you, and the one question that separates a prompt library from a written design specification.

What NightMarquee gives you#

The catalogue leans to mobile and ecommerce, with smaller runs of SaaS, landing, hero, portfolio and 3D pieces. The one-off tier is worth the arithmetic before you pick a plan: $59 paid once is less than five months of the $12 subscription.

The four variants, and what changes between them#

Nearly everything in this category claims to work with every tool. NightMarquee does something more specific than claim it.

The four files are different documents, not one text with a tool name swapped at the top. The Lovable variant is written as React component instructions and the Cursor variant as a project file map, and they diverge from the first line rather than at the edges.

You can confirm that for nothing:

1. Pick one of the six free prompts.
2. Open all four of its variants.
3. Read the first screen of each.

Five minutes, no card, and you will know whether per-tool rewriting is a thing you value or a thing you were sold.

Where the prompt gets pasted#

A prompt is only a file until it reaches the tool you build in, and there are two routes here. Copy from the site, or install their MCP server and have prompts arrive as a command inside Claude Code or Cursor, with a sign-in that unlocks the paid ones in place.

The second route is the one that changes a working day. A prompt you have to go and fetch gets fetched once. A prompt that lives in the editor gets used on the third revision too, which is where most of the design work actually happens.

It is also where the four variants earn their keep or fail to. In Cursor you want the file map. In Lovable you want component instructions, on top of what Lovable's own design layer already decides for you. A single prompt written for neither is one you translate by hand before the model sees it, every time you open it.

Scroll the outcome before you spend a token#

Every prompt has a demo page, and the demos are hand built pages served on the same domain rather than a recorded video in a frame. They scroll. One runs a particle canvas; another scrubs a video against scroll position instead of autoplaying it.

In a category that sells text you cannot read, scrolling the result first is a real advantage, and it makes the decision cheap in a way a screenshot never does.

What the demo has that the prompt does not#

Every demo page loads production video from a media host. Read the free prompts and most of them never mention video at all, and two ask for a single self-contained file with no external assets. Their demos load one anyway.

The film layer in the preview is an asset class the prompt does not request. So the output will differ from the preview on that axis whatever you paste it into, and closing the gap is yours: source the footage, or build the section without it.

Four steps, five minutes, no money:

1. Open a free prompt and read it to the end.
2. Open its demo page beside it.
3. List everything in the demo the text never asks for.
4. That list is your afternoon.

This is not a flaw peculiar to one seller. It is the default state of a category that sells a written artefact and advertises it with a finished page, and it is worth knowing how to spot before you buy anywhere.

The question underneath all of it#

Catalogue size is the easiest thing to compare and the least useful. The thing worth comparing between prompt libraries is how much is decided inside a single file.

A prompt of a few hundred words carries a concept and a stack. It hands the model a direction and lets the model resolve everything under it, which is why ten people working from ten short briefs produce ten pages that rhyme.

A specification decides instead. Thanor writes each one to a fixed shape:

• A palette of four to six hex values, each with a named role, and the neutral's hue bias stated, because a pure mid-grey is the clearest tell of a generated design.
• Two named faces with the fallback stack written out, a scale in px, the weight at each size, line heights and letter-spacing.
• A grid, a gutter, page margins and a vertical rhythm as real numbers, in the shape of 160 / 96 / 48.
• Named curves with real values such as cubic-bezier(.16,1,.3,1), durations in ms, and which moves run continuously off scroll progress rather than firing once.
• Numbered sections, each buildable without having read the one before it.
• A never list written for that design, not generic taste advice.
• Acceptance checks a person can run by eye, in the shape of the headline resolves through a mask, not a fade.

Every file clears a 1,200 word floor before it enters the library, and an automatic gate rejects it for a missing block, a colour count under four, or motion described with no curve and no duration.

If you are choosing today#

The first three are one product. The fourth is a different one, and which you need depends on whether you already know what the page should look like.

If you do, buy the effect and move on. If you do not, no quantity of short prompts will supply it, because the part you are missing was never in them.

Go and look at what Thanor writes into one, and judge it the way this page has asked you to judge everything else: by what it puts in writing.