Kimi K3 Designs Websites That Feel Like MOVIES For Just $1

At a glance

Nick Saraev builds three cinematic scroll driven landing pages end to end and shows every keystroke: the ideation prompt, the prompt that turns a concept into a video prompt, the video model settings, the frame interpolation pass, the site build, and the deploy. The mechanic behind these sites is simple once he says it out loud. There is no elaborate 3D scene. There is one short macro video in the background, every frame of it becomes a scene, and the frames are bound to scroll position, so the page moves forward when you scroll and backward when you scroll up. Everything else is scaffolding on top of that. The tool chain is Kimi K3 from Moonshot AI as the brain and the coder, Higgsfield Cinema Studio 3.0 as the video generator, a ByteDance AIGC model for frame interpolation from 30 to 60 FPS, and Netlify for hosting. Then he does the thing that makes the video worth watching twice: he stops driving the tools by hand, installs the Higgsfield MCP server into Kimi Code, and collapses the whole chain into a single prompt that goes from "you go nuts" to a deployed website. His gallery of the results is live at kimik3designs.netlify.app.

What is actually going on in one of these sites

He opens on a finished page and scrolls it up and down. The thing you cannot tell from a screen recording is that he is scrolling, not playing anything. There is a sense of forward motion through a video, and the natural assumption is that the website itself is what moves and rotates.

It is not. At its core, he says, you take every frame in a video, turn them into what are effectively separate scenes, and bind them to a scroll animation, so scrolling one way moves you forward through the video and scrolling the other way moves you back. The page is stationary. The film is what advances. Down in the corner there is also a tour control, a flow button that walks a visitor through the whole thing on auto play if they would rather not scroll it themselves.

Then he makes the point that reframes the entire build: the creative is the heavy lifter. Despite how much motion the finished site seems to have, the motion is a video in the background. The interactive layer matters too, but that is the part Kimi can do easily. The part that decides whether the site looks like a million dollars or like a template is the eight seconds of macro footage you feed it.

The four jobs the pipeline has to do

Before he touches anything, he lists what has to happen. It is worth keeping in front of you because every later step is one of these four.

1. Generate the creative asset, the background video itself.
2. Process that video so it is genuinely high quality, the macro, zoom in journey look rather than a generic clip.
3. Interpolate frames so the scroll is smooth instead of chunky. This is the step he says most people skip, and skipping it is why other people's versions of this look bad.
4. Have Kimi K3 wrap it in a nice bow and present it as a working site.

Getting Kimi K3 and Kimi Code running

The model is Kimi K3, which he introduces as the new frontier model from Moonshot AI, Kimi's creators. His read on it: K3 is really solid and sits on par with Fable, GPT 5.6, and the other frontier models of the moment. He says hello to it in the chat interface and immediately points out the limitation. A chatbot cannot build you a website. You need the coding agent.

So you go to Kimi Code, not Kimi chat. Two practical warnings come with it. First, subscriptions are heavily throttled and bottlenecked at the moment, precisely because this is such a cheap way of producing high quality output, so you may not be able to sign up the second you try. Second, when you can, the plans are named after musical tempos: Moderato, Allegretto, Allegro, and Vivace. Sign up is one click through to a Stripe page.

Installing the CLI is a single copy and paste. Higgsfield and Moonshot both hand you the line; he copies it from the Kimi Code page and pastes it into a terminal.

# the install line copied from the Kimi Code page (a curl -fsSL … | bash one liner)
# he is running it here as an upgrade, since he already had it installed
curl -fsSL <the install URL Kimi gives you> | bash

Then three commands are all you need to be operational:

kimi          # launch Kimi Code
/login        # OAuth into your Kimi account, opens a browser page if you are not signed in
/clear        # wipe the context and start clean before every new job

He is candid that a terminal looks intimidating if you are new, and equally candid that you do not need to understand curl -fsSL. You paste the line, Kimi does the rest.

He also flags where the prompts live: everything he uses is published free in his community, Maker Zero, linked at the bottom of every one of his videos. The prompts below are what he reads on screen.

The ideation prompt: let the model ideate, you supply the taste

The first prompt is a concept generator. He is explicit about which part of it does the work.

You're the sole creative director and senior front-end author for a premium
interactive portfolio.

Give me creative ideas videos for 10 radically different scroll-driven
commercial landing pages.

The video must involve a single continuous 8-second extreme macro journey.
This journey is what we use as the main captivating feature of our websites,
so make this as interesting as possible.

"They need to be scroll driven," he says, and "a single continuous 8 second extreme macro journey." Those two constraints are the whole brief. Everything else, the industry, the subject, the palette, is left open on purpose, because the point is volume.

Then comes the argument underneath the technique, and it is the most quotable idea in the video. For humans, coming up with ideas is computationally difficult. For an AI, which he describes as seeing ideas as points plotted on a big graph of possible concepts, ideation is easy. What humans are still great at is taste, picking the thing that will actually land with an audience. So the division of labor is: have the model ideate across a very large space of possible solutions, then do the one thing you are better at, which is pick your favorite.

He does not need the ten concepts to be the best concepts in the world. He needs ten to choose three from.

Reading the concepts out loud

Two of the ten get read verbatim on screen, and they are worth reproducing because they set the level of specificity a video model needs.

The Pour, luxury whiskey.

"We start buried inside a single bubble in the whiskey. The sphere's skin glows amber around the lens. Then the camera pushes through the membrane into liquid gold, slides past rising carbonation threads, climbs the meniscus of the glass, and breaks the surface tension."

His verdict: interesting, but he wants something more impressive than a luxury whiskey site.

The Capsule, skin care and pharma.

"Begin on the surface of a dissolving capsule. The gelatin wall is melting in front of the lens like slow ice. The camera pushes through the membrane into a cloud of white macro powder, drifts through the granules, each one catching light like a snow globe, and condenses onto a single droplet that falls and lands on skin. The pore it touches blooming smooth."

That one he keeps. Notice the shape both share: start buried inside something at microscopic scale, push through a membrane, travel, and break out. That is the "extreme macro journey" pattern the ideation prompt asked for, and it is what makes the scroll feel like forward motion rather than a pan.

He copies the winners into a Google Doc titled concepts. Then a small operational decision that pays off later: rather than generate one concept, he takes three, because generating three costs the same wall clock time as generating one, and he will keep whichever he likes best.

At this point he makes the claim that reorders the whole job: once you have the concepts, you are about 50% of the way to a jaw dropping experience. Not because the code is trivial, but because the creative is what carries the site, and the code is what Kimi handles easily.

The second prompt: turn a concept into a video prompt

Concepts are for humans. Video models want a different kind of text, so he clears Kimi and runs a translator prompt. The key move is teaching by example: he pastes in a sample video prompt whose level of detail he wants matched.

Use the below concept to generate a high-quality macro video prompt for an
advanced video model.

For example:
"A single continuous 8-second cinematic engineering shot moving forward from
above a stormy cold North Atlantic Ocean…"

<paste one concept from the concepts doc here>

"What I'm doing here," he says, "is I'm giving it an example of the level of detail that I would like." He is not describing the style he wants in the abstract. He is showing a sample that he already knows performs well and asking for the same shape.

The output he gets back opens like this:

"A single continuous 10 second cinematic macro shot opening on the hexagonal lattice of a single ice crystal at molecular scale…"

Roughly the same length as the example, roughly the same density. That is the tell that the by example approach worked. Two prompts in, and he has gone from nothing to a production ready video prompt without writing a sentence of creative himself.

The 30 FPS problem, and why interpolation is not optional

He has tested a few dozen video models and approaches, and this is the combination he landed on. Today's generator is Cinema Studio 3.0, on Higgsfield. He is clear it is not the only option, he just wanted the best available at the moment.

Then the catch that most people building these sites never address. Cinema Studio output comes back at roughly 30 FPS. On a normal video that is fine. On a scroll heavy cinematic journey site it is not, because the scroll is what advances the frames, and at 30 FPS you can see every individual frame tick past as you scroll. It reads as choppy. He calls it a very bad experience for the user.

The fix is frame interpolation, and he draws it on screen. Take frame 1 and frame 2, and generate a new frame that belongs between them. Take frame 2 and frame 3 and do the same. An element sliding across the screen now has twice as many positions to occupy, so the motion is continuous instead of stepped. He runs it through a ByteDance AIGC model to go from 30 to 60, and his claim is blunt: once you are at double the frame rate, the vast majority of all scrolling issues disappear.

Inside Higgsfield: the exact settings

Higgsfield is where the film gets made. He describes it as a big video, audio, and prompt library that contains more or less everything you need to copy and paste into a model like Kimi K3 to make this work. He has an account already and skips past the current signup promo.

Navigation: explore at the top, then video, then Cinema Studio 4K. Inside, the version picker offers 3.5, 3.0, and 2.5. He picks 3.0 on purpose, because he does not need the extra scene change features that the newer version adds. Then he pastes in the video prompt Kimi just wrote.

The settings, and his reasoning for each:

• Aspect ratio 16 by 9. Desktop experience. If you are building for mobile, or generating creative for Instagram, click 9 by 16 instead, which he notes more people are doing lately.
• Resolution 1080p. He does not want the generation to take too long.
• Duration 8 seconds. The slider also offers 10 and 16. Eight makes the most sense for this, which is why the ideation prompt asked for 8 seconds in the first place.
• Cost about 80 credits for that combination. He does not want to burn credits carelessly, so he checks it before sending.
• Mode. Custom, multi shot, smart and so on are available. He leaves the default.

He hits generate. The prompt goes in on multiple lines, which he notices looks ugly and hopes will not affect the generation. A modal opens saying "new project processing," and from there the only job is to wait.

Except he does not wait. He batches. He goes back to Kimi, asks it to do the other two concepts as well, and feeds the resulting prompts into two more Cinema Studio generations. Now three are rendering at once, on the same settings, 16 by 9.

The reasoning is explicit and it generalizes far past this video: one generation takes X amount of time, and the service can work on multiple things at once, so waiting serially is pure waste. He is trading money in credits for time and efficiency. And there is a second reason that is really the honest one:

"I'm never really 100% of the time happy with the outputs of AI. I'm satisfied maybe 30 to 50% of the time, which means typically I need more than one sample or more than one output in my batch in order to be happy."

Batch size is a hit rate calculation. If one in three generations is good enough, generate three.

The results come back. The capsule dissolving and opening, which he calls really high quality and very clean, micellar water territory. The city walk through, which turns out to be a 3D city that is also a circuit on a silicon wafer. He can scrub them in Higgsfield's editor, but scrubbing in an editor is not the point. He wants them tied to a scroll wheel. He takes a break, comes back to three finished renders, and clicks the download button on each.

Stop clicking: wire Higgsfield into Kimi with MCP

Here is where the video turns from a tutorial into a workflow argument. "Doing this stuff manually is starting to get pretty tiring," he says, and he is right, because he has just spent five minutes as a human clipboard between two browser tabs.

The fix is MCP. On Higgsfield, go to MCP and CLI, and click the copy button. That gives you a config block that defines a connector to Higgsfield's back end. Paste it into Kimi and say four words:

Install this MCP.

Kimi goes through and creates the connector behind the scenes, writing an mcp.json file onto your computer. That is just how MCP config works. Then:

/clear          # this initiates a new server, so the MCP is live in the session
/mcp-config     # only if you are not already authenticated: walks you through
                # a short sign up wizard back to Higgsfield to sign in

What that buys him: instead of asking Kimi to write a prompt that he then copies into the Higgsfield UI, he can now tell Kimi to generate the Higgsfield videos directly. Same for upscaling. He can ask for a pipeline where every generated video is automatically upscaled from the start. Two stages of the pipeline stop being a browser tab.

The first thing he does with it is the interpolation pass. He grabs the three downloaded filenames with a copy, pastes them in, and writes one sentence:

Interpolate this with ByteDance AIGC.

Kimi finds the right Higgsfield model and runs it. The instruction underneath is simply "double the frames."

Two Kimi instances, and the YOLO flag

Kimi is intelligent, he says, but it can also be pretty slow at times. His answer is not to wait, it is to parallelize himself. While the left hand instance is firing Higgsfield requests and improving the media, he opens a second instance to start scaffolding the website.

kimi --yolo

He describes --yolo as a funny mode that prevents the agent from ever having to ask him a question. You only live once mode. Just let the agent do everything. It is the same bargain every agentic CLI offers: you trade the approval prompts for speed, on the assumption you are working in a directory you are willing to let it touch.

The site prompt

Now the build. He heads to his site prompt, a long block he keeps in Maker Zero, copies the whole thing, and pastes it into Kimi K3. He is upfront that it is long and that he is not reading it out, because it is available free.

What matters is the interface it exposes, because that tells you what a site prompt of your own has to accept. It asks for two inputs:

1. The video file, by path on your computer. The prompt literally shows an example, capsule.mp4. He opens Finder, goes to downloads, and hands it the capsule clip.
2. The concept prompt that was used to generate that video. He goes back to Higgsfield, clicks the generation, and copies out the exact prompt he fed in, then pastes it in.

That second input is the clever part. The site prompt does not just get pixels, it gets the creative intent behind the pixels: what the journey is, what the subject is, what the mood is. That is what lets the copy, the section structure, and the color treatment match the film rather than fight it. From there the prompt runs a long, in depth series of steps to produce the motion graphic design.

And then he says the thing he has clearly been waiting to say: "You're probably already seeing areas that we can optimize the hell out of this," because every one of these steps was made manually. The pipeline is about to get woven together.

The finished site

The build lands. He scrolls through it: a complicated universe of circuit boards and, at the end, a burst of light. The page has real interactive furniture on it, not just a moving background:

• A hold to reserve call to action, the kind of thing you would use for a product pre launch.
• A watch it dissolve control that jumps you from one section to the next.
• Manual scroll at whatever speed you want.
• A flow tour button in the bottom corner that plays the whole thing through as a video for visitors who do not want to scroll.

It is, in his words, a live and working example. And that is end to end.

Publishing it: Netlify in about ninety seconds

Hosting is a static site problem, so any static host works. He uses Netlify, and notes Vercel works exactly the same way. His own gallery of every site from this method lives at kimik3designs.netlify.app.

The flow, which is the same on most platforms:

1. Log in to the host.
2. Go to user settings.
3. Under applications, find OAuth.
4. Click new access token. He names it something like "example token" and sets a short expiry, today or tomorrow.
5. Generate token, copy it.
6. Paste the token into Kimi and hand it over.

Kimi takes the key and creates the site on his Netlify account. Within moments it is 100% online at a netlify.app address, flow tour button and all, and anyone can open it.

Two things worth noticing here that he does not spell out. First, an agent with a deploy token can deploy, which is exactly as powerful and exactly as sharp as it sounds, and it is why he generates a short lived token rather than a permanent one. Second, this is the only step in the entire pipeline that touches credentials, which tells you how thin the actual infrastructure is: it is a folder of static files.

The motherload: one prompt, end to end

"I told you this is going to be the motherload." One prompt containing every step above, start to finish. You paste the whole thing into Kimi and it takes over.

It asks exactly two questions:

1. What subjects, products, industries, materials, or ideas are you interested in?
2. Do you have any creative direction references or constraints, or should I grab everything?

His answers, both times: "You go nuts."

That is the entire input. Concept ideation, concept selection, video prompt authoring, Higgsfield generation, interpolation, site scaffolding, and deploy all run underneath. He notes the terminal looks slightly different in this segment simply because he switched terminals, not because anything else changed.

The gallery: what "go nuts" produces

The end results are the sites he opened with. He tours them:

• The hot air balloon. You rise into the sky until the whole horizon opens up, and you are legitimately inside the balloon looking out.
• The cave. You scroll through a dark, dank cave and arrive at what looks like a record player whose head is on fire. He admits this one is a little more zany.
• The foundry. Something producing brass, which resolves into what looks like an old school printing press.
• The circuit city, the one he built live, a 3D city that is actually a circuit on a wafer.
• The capsule, the skin care concept, dissolving and blooming.

How many you make comes down to two budgets and nothing else: how much time you want to spend, and how many credits.

He closes on the offers. Every asset and every prompt is free in Maker Zero. Maker School, the first link in his description, is his 90 day accountability program for people who want to sell these interactive experiences rather than just build them, with a 90 day full money back guarantee if you do not land client number one. Then, verbatim: "Anyway, I'll stop shilling my own products."

The reproduction spec: build one yourself

Everything below is drawn from what the video does on screen. Where he pastes a prompt he keeps in Maker Zero without reading it, this section rebuilds a working equivalent from the behavior shown, and says so plainly rather than pretending to quote him.

Step 0: tools

Step 1: install and authenticate Kimi Code

curl -fsSL <install URL from the Kimi Code page> | bash
kimi
/login          # OAuth in a browser
/clear          # always start a job from a clean context

Step 2: ideate concepts (verbatim from the video)

You're the sole creative director and senior front-end author for a premium
interactive portfolio.

Give me creative ideas videos for 10 radically different scroll-driven
commercial landing pages.

The video must involve a single continuous 8-second extreme macro journey.
This journey is what we use as the main captivating feature of our websites,
so make this as interesting as possible.

Read all ten. Keep three. Paste them into a doc. Your only job in this step is taste.

Step 3: turn a concept into a video prompt (verbatim from the video)

Use the below concept to generate a high-quality macro video prompt for an
advanced video model.

For example:
"A single continuous 8-second cinematic engineering shot moving forward from
above a stormy cold North Atlantic Ocean…"

<paste one concept here>

The example line is doing the work. Swap it for any prompt you already know performs well and the model will match its density.

Step 4: render the film

In Higgsfield: explore → video → Cinema Studio 4K → version 3.0. Paste the prompt from Step 3, then set:

• Aspect ratio 16:9 for desktop, 9:16 for mobile or Instagram creative
• Resolution 1080p
• Duration 8 seconds
• Confirm the credit cost, which was about 80 on that combination
• Leave the mode on the default

Generate. Then immediately queue your other two concepts so three render at once. Download each when it finishes.

Step 5: wire the Higgsfield MCP into Kimi

On Higgsfield, open MCP and CLI and copy the config block. In Kimi:

<paste the Higgsfield MCP config block>

Install this MCP.

Kimi writes mcp.json. Then:

/clear          # initiates a new server so the connector is live
/mcp-config     # only if you are not already authenticated to Higgsfield

From this point Kimi can generate and upscale Higgsfield videos directly, with no browser tab in the loop.

Step 6: interpolate to 60 FPS

Interpolate this with ByteDance AIGC.

capsule.mp4
circuit.mp4
crystal.mp4

Doubling the frame rate is the whole objective. Anything that reliably takes 30 to 60 will do.

Step 7: build the site

Open a second agent so rendering and building overlap:

kimi --yolo

Then paste your site prompt. It must accept two inputs, because those are the two the video's prompt asks for: the path to the interpolated video file (for example capsule.mp4) and the original concept prompt used to generate it. A working equivalent, rebuilt from what the finished sites do rather than copied from his text:

You are a senior front-end author. Build a single self-contained cinematic
scroll-driven landing page from the assets below. Ask me for anything missing
before writing code.

INPUTS
- Video file: <path to the interpolated 60 FPS clip, e.g. capsule.mp4>
- Concept prompt used to generate it: <paste the exact Higgsfield prompt>

BUILD
1. Extract the clip to a numbered still-frame sequence, one image per file.
2. Draw the current frame onto a full-viewport canvas fixed behind the content.
3. Map scroll progress (0 at the top of the page, 1 at the bottom) to a frame
   index, and ease the current frame toward the target so motion feels weighted.
   Scrolling up must run the sequence backward, identically.
4. Stop drawing entirely once scrolling stops. Redraw only when the frame changes.
5. Derive the section structure, headline copy, and palette from the concept
   prompt, so the words on screen match the journey happening behind them.
6. Add: a primary call-to-action pinned to the climax of the journey, a
   section-to-section jump control, and a "flow tour" button that auto-plays the
   whole journey for visitors who do not want to scroll.
7. Respect prefers-reduced-motion. Animate transform and opacity only.

Output one self-contained folder I can drop on a static host.

Step 8: deploy

On Netlify: user settings → applications → OAuth → new access token, set a short expiry, generate, copy. Then hand it to Kimi:

Deploy this site to Netlify. Here is a short-lived access token: <token>

Vercel or any static host works the same way. Revoke the token when the deploy is done.

Step 9: collapse it

Once you have run the whole thing once by hand, write the version of the pipeline that runs itself. His asks two questions and nothing else:

1. What subjects, products, industries, materials, or ideas are you interested in?
2. Do you have any creative direction references or constraints, or should I grab
   everything?

That is the real lesson of the last five minutes of the video. Do it manually, watch where you are acting as a clipboard, then delete yourself from those steps.

An honest footnote: what the dollar actually covers

The title says a dollar. The first sentence of the video says "literally just a dollar or two." Here is every cost figure the video actually states, and every one it does not.

Read fairly, the dollar or two is a claim about the model, not the pipeline. Kimi K3 doing the ideation, the prompt translation, and the site build genuinely is inexpensive, and that is the part the video is really about. The video generation is a separate budget, priced in Higgsfield credits, on top of a Higgsfield subscription that also is not priced here. If you show up with a hard one dollar limit you will get the website and not the film.

Two other things worth flagging, in the same spirit. His own hit rate number, satisfied with AI output 30 to 50% of the time, means the honest cost is two to three generations per usable clip, not one, and he builds that into his workflow rather than hiding it. And the site prompt and the end to end prompt, the two artifacts that do the heaviest lifting, are not read on screen. They are free, but they live behind a signup at Maker Zero. The technique is fully reproducible without them, which is what the spec above is for, but the video is a demo of prompts you go elsewhere to collect.

None of that dents the core idea, which is genuinely good and generalizes past cinematic websites: let the model ideate across a wide space, spend your own judgment only on selection, batch anything that renders in parallel, then wire the tools together with MCP so the parts of the job where you were acting as a clipboard disappear.

Key takeaways

• The site is not moving, the film is. Every frame of a short macro video becomes a scene bound to scroll position. Scroll forward, the video advances; scroll back, it reverses. That is the entire mechanic.
• The creative is the heavy lifter. The interactive layer is the easy part for a model like Kimi K3. The eight seconds of footage decide whether the page looks expensive.
• Ideation is cheap for AI, taste is cheap for you. Ask for ten radically different concepts, then do the one job you are better at: pick the one that will land.
• Constrain the brief, not the content. Two constraints did all the work in the ideation prompt: scroll driven, and a single continuous 8 second extreme macro journey.
• Teach style by example. The concept to video prompt step works because he pastes in a sample prompt whose detail level he wants matched, instead of describing it.
• Interpolate, or it looks cheap. Cinema Studio output is 30 FPS, and on a scroll bound site you see every frame tick. Doubling to 60 with a ByteDance AIGC pass makes most scroll problems vanish. He says this is the step most people skip.
• Batch everything that renders in parallel. Three generations take about as long as one. He is satisfied with AI output only 30 to 50% of the time, so batching is not extravagance, it is the hit rate.
• Give the site builder the concept, not just the video. The site prompt takes the film path and the original creative prompt, so the copy and structure match the journey.
• MCP removes you from the loop. Installing the Higgsfield MCP into Kimi turns two browser tab stages into sentences, and kimi --yolo in a second instance lets rendering and building overlap.
• Then collapse the whole thing into one prompt. The final act, and the real lesson: run the pipeline manually once, find where you were the clipboard, and write the prompt that does all of it.

Chapters

0:00 The claim: cinematic websites for a dollar or two 0:31 What is actually happening: frames bound to scroll 1:03 The four jobs the pipeline has to do 1:46 Kimi K3 and Moonshot, chat versus Kimi Code 2:29 Plans, throttled signups, and the install command 3:24 kimi, /login, /clear 3:53 The ideation prompt: 10 scroll driven concepts 4:36 Why AI should ideate and you should pick 5:06 The Pour, and the Capsule concept that wins 6:07 Concepts done means you are 50% of the way there 6:53 Turning a concept into a video prompt, by example 8:04 Cinema Studio 3.0 and the 30 FPS problem 8:45 Frame interpolation explained, 30 to 60 with ByteDance AIGC 9:43 Inside Higgsfield: explore, video, Cinema Studio 4K 10:40 The settings: 16 by 9, 1080p, 8 seconds, about 80 credits 11:50 Batching three generations at once 13:02 His hit rate: satisfied 30 to 50% of the time 13:29 The renders land, download the clips 13:43 Install the Higgsfield MCP into Kimi 14:53 /clear, /mcp-config, and mcp.json 15:22 "Interpolate this with ByteDance AIGC" 16:06 A second instance with kimi --yolo 16:21 The site prompt: video file plus concept prompt 17:35 The finished site, hold to reserve and flow tour 18:18 Publishing to Netlify with an OAuth token 19:16 The motherload: one prompt, end to end 19:59 "You go nuts" twice, and the results 20:43 Maker Zero, Maker School, and the sign off

Notable quotes

"Kimi K3 can design absolutely gorgeous, beautiful websites just like these for literally just a dollar or two." (0:00)

"At its core, what we're doing is we are taking every frame in a video, we're turning them into different scenes, essentially, and then we're binding them to some sort of scroll animation such that when I scroll up, I go forward, and when I scroll down, I go back." (0:40)

"We also need to interpolate frames, do a couple of other things so that it's not chunky and it's smooth, unlike what a lot of other people do." (1:24)

"This concept is one that you can come up with yourself, but in our AI era, I much prefer to just have AI do the concept generation for me." (1:48)

"The cool part about AI is for humans, coming up with ideas is very computationally difficult. For AI, which views ideas as sort of these plots on a big graph of possible concepts, ideation's actually really easy." (4:42)

"The cool thing that humans are great at right now is employing our taste to pick things that we think might actually sound good or make sense to our audience." (4:52)

"Once we have the concepts, we're actually already about 50% of the way there to designing a smooth, sort of jaw dropping experience like this." (6:14)

"The creative is really the heavy lifter of this site." (6:37)

"What I'm doing here is I'm giving it an example of the level of detail that I would like." (7:12)

"The issue with 30 FPS on one of these like scroll heavy, you know, cinematic journey sites is you can actually see every individual frame as you scroll through, and so it's really choppy." (8:33)

"I'm going to be using ByteDance AIGC to take us from 30 to 60, and then once we're at double the FPS, the vast majority of all scrolling issues tend to disappear." (9:29)

"I'm never really 100% of the time with the outputs of AI. I'm satisfied maybe 30 to 50% of the time, which means typically I need more than one sample or more than one output in my batch in order to be happy." (13:02)

"That said, doing this stuff manually is starting to get pretty tiring." (13:47)

"This is just sort of like a funny mode that prevents it from ever having to ask me a question. It's on like the you only live once mode, so just let the agent do everything." (16:10)

"You're probably already seeing areas that we can optimize the hell out of this. And the reason for that is because what we've done here is we've manually sort of made all these steps." (17:24)

"I told you this is going to be the motherload. This is a single prompt that contains every one of these steps end to end." (19:22)

"It really just boils down to how much time you want to spend and how many credits." (20:45)

Resources mentioned

• Kimi, the model and the chat interface. Kimi Code is the coding agent used throughout, installed from a curl one liner, with plans named Moderato, Allegretto, Allegro, and Vivace.
• Moonshot AI, the lab behind Kimi and the K3 frontier model.
• Higgsfield, the video, audio, and prompt library where every clip is generated. Cinema Studio 4K, version 3.0, at 16 by 9, 1080p, 8 seconds, about 80 credits per generation. Its MCP and CLI page provides the connector config.
• ByteDance, whose AIGC model runs the 30 to 60 FPS frame interpolation pass through Higgsfield.
• Model Context Protocol, the standard behind the mcp.json connector that lets Kimi call Higgsfield directly.
• Netlify, the static host used for the deploy, via an OAuth access token generated under user settings and applications.
• Vercel, named as an equally valid static host.
• kimik3designs.netlify.app, his live gallery of every cinematic site produced with this method.
• Stripe, the checkout the Kimi subscription flow hands off to.
• Maker Zero, his free community, where every prompt in the video is published including the site prompt and the single end to end prompt.
• Maker School, his 90 day accountability program for selling these builds, with a 90 day money back guarantee tied to landing client number one.
• Alternative models he says the prompts work with: ChatGPT, Claude, or any other frontier model.
• Nick Saraev on YouTube, the creator's channel.