The Evolution of AI Video Generation, and a Video Claude Made From One Prompt

Four years ago, an AI-generated video was a few seconds of melting, silent blur that impressed researchers and nobody else. Today a marketing team can type a paragraph and get a 4K clip with sound. And in the same period, the best-known video model in the world launched, went viral and was shut down.

This post traces how AI video generation evolved, explains the two very different ways AI now makes video, and shows a complete promo video that Claude produced from one prompt. The prompt is included so you can try it yourself.

From research demo to 4K with sound: a timeline

Three things improved, roughly in this order: coherence (objects that stay the same object from frame to frame), control (camera moves, first and last frames, consistent characters) and sound. What has not improved as much is text. Generative models still draw letters as pixels, which is why signs, captions and logos in AI footage so often come out garbled.

Two ways AI makes video

This is the distinction most articles miss, and the one that decides which tool you should use.

Generative video: painting pixels

Veo, Gemini Omni, Kling, Runway and (until this year) Sora are diffusion-style models. They start from noise and paint every frame, guided by your prompt. They are extraordinary at realistic scenes: a drone shot over Phewa Lake, a chef plating momos, rain on a tea garden.

Their weaknesses are the flip side of that. You cannot edit one element without regenerating the clip, results vary from run to run, clips are short (typically around eight to ten seconds before extension), and text and logos are unreliable.

Code-generated video: writing the animation

Claude works differently. It does not generate video pixels at all. It writes code that draws the video: an HTML canvas animation, an SVG sequence or a Remotion project built in React. A renderer then plays that code frame by frame and saves it as an MP4. Claude Code can run that rendering step itself.

The result looks like motion graphics rather than filmed footage. In exchange you get things generative models cannot offer:

• Text is always perfect, including Devanagari, because it is rendered with a real font.
• Brand colours and fonts are exact, not approximated.
• Every element is editable. Change the phone number, the dates or the headline and re-render; nothing else shifts.
• It scales. One template can produce fifty variants, one per product, route or branch, each consistent with the others.

A video Claude made from one prompt

To show what this looks like in practice, I gave Claude the prompt below and nothing else: no images, no footage, no template. The business is fictional.

Make a 15-second, 16:9 animated promo video for a fictional Pokhara trekking company called Lakeside Treks.

Scene 1 (0–4s): dawn over layered Himalayan ridges with a fishtail-shaped peak, the sun rising behind them, slow parallax, birds crossing, and the ridges reflected in a lake at the bottom. The title "Lakeside Treks" rises in, with "Pokhara · Annapurna · Everest" underneath.

Scene 2 (4–9s): three benefit lines slide in one after another, each with a tick: "Licensed local guides", "Permits handled for you", "Small groups, fair prices".

Scene 3 (9–13s): a phone slides up showing a WhatsApp-style chat. A customer asks whether Annapurna Base Camp is open in November; typing dots; an instant reply with season, two departure dates and trip length; the customer replies "Yes please!". Beside the phone: "Ask anything. Get an answer in seconds."

Scene 4 (13–15s): end card with a simple mountain logo, the name, a green "Book on WhatsApp" button and a placeholder phone number.

Warm dawn palette, clean sans-serif type, smooth easing, no stock footage or images. Write it as a single HTML file with a canvas, where every frame is a pure function of time, so it can be rendered frame by frame to MP4.

And this is the result:

How it was made

1. Claude wrote the animation: a single HTML file of about 350 lines that draws each frame on a canvas: the sky gradient, the sun, four parallax ridges, the reflection, the birds, the text and the phone chat.
2. Claude Code rendered it: it opened the file in a headless browser, captured all 450 frames (15 seconds at 30 frames per second) and encoded them with ffmpeg into an H.264 MP4.
3. Review and two rounds of fixes. The first render had the sun sitting behind the title, a half-empty phone screen and chat text too small to read on a phone. I described the problems in plain English, Claude changed the code, and the video was re-rendered. That review step is the part a human should never skip.

From prompt to final file took minutes, not days. And because the video is code, making a version for a different trek, with Nepali text or a real phone number, is a one-line change and a re-render.

What this approach is not good for

To be clear about the limits: this approach will not give you a realistic shot of trekkers on a trail, a guide's face, or a hotel room. For those, use a generative model, or better, a real photographer. Code-generated video is for explainers, announcements, offers, data stories, social posts and templated content, which is a large share of what small businesses actually post.

Which approach for which job

What this means for businesses in Nepal

For tourism, hospitality, education and retail, the businesses that post the most video in Nepal, the cost of a decent short promo has dropped from a production day to an afternoon. That does not mean every business should flood social media with AI clips; audiences already scroll past generic ones. It means the businesses that combine real photography of real places with fast, well-made AI motion graphics can post consistently without a production budget.

The skill that matters is no longer editing software. It is writing a clear brief, the same skill as writing a good prompt, and having the judgement to reject a result that is not good enough.

How I can help

I help businesses use AI video in three ways:

• Branded templates. A set of code-generated video templates in your colours and fonts, for offers, announcements and seasonal campaigns, as part of a digital marketing engagement.
• Video at scale. Automations that turn a spreadsheet of products, routes or courses into a finished video for each, using the template approach shown above.
• Training your team. A practical session for marketing and communications staff on briefing generative tools and Claude, reviewing output and staying on the right side of copyright and disclosure. It is the creative-team version of my AI training.

If you want a video like the one above for your own business, or want your team to be able to make them, send me a message with what you would like to promote.