A few weeks ago I came across a post on X about making $10k-looking launch videos with Claude. The whole thing was made through a conversation and some code, without a video editor or a motion designer anywhere near it.
I tried it that same day. Since then, X has pretty much flooded with people doing exactly this, and Product Hunt is next. Every launch will soon have a video that looks like a studio made it.
I'm not a motion designer, and I don't know After Effects. In the last few weeks I've made launch videos, product demos, a silent loop for a landing page and a set of ads, all with Claude Code. The launch video for my marketing tools site mktbox, which went out with its Product Hunt launch, and the "Made with Nocta" loop for Nocta, my product shot app, are two of them. Here's exactly how it works, including what went wrong along the way.
The idea: a video is a web page that moves
This is the part that made it click for me.
Each video starts as a web page. Every element on it, a headline, a cursor, a product screenshot, knows where it should be at any given moment, down to the hundredth of a second. A script then opens that page and takes a picture of it 60 times per second, like a flipbook, and stitches the pictures into a video file.
Because it's a page, Claude can write all of it. The animation, the camera moves, the timing, the colours, your fonts, your logo. You never touch a timeline.
The process, step by step
Every video goes through the same six steps. Start to finish, one takes me about an hour and a half. If I want changes after the first version, two to two and a half hours.
1. The brief. Before anything gets built, Claude asks me questions. What's the video for? Who sees it? Where does it play: a landing page, LinkedIn, Product Hunt? How fast should it feel? What should it look like? Sound or silent?
2. The script and the beat map. Claude writes the script, then a beat map: a table of what happens second by second. For the mktbox video, it looked like this: from 0 to 5 seconds the logo lands and the headline appears, from 5 to 11 the tools menu opens and the list ripples in, from 11 to 19 one tool grows into its own page. You approve this before a single frame moves.
3. Still frames. Claude shows me key moments as images. It's much faster to fix a layout on a still than in a video.
4. A fast draft. A lower quality version renders in a few minutes, so I can watch the whole thing and give notes.
5. Notes and fixes. "Slower here." "That text needs more time." "The logo should land on the beat."
6. The final render. Full quality, only once the draft is right.
Sound, without paying for it or breaking a licence
Music was a rabbit hole. We kept checking sources until we had a short list where everything is free for commercial use with no credit line: Mixkit and Pixabay for music and sound effects, Freesound but only files marked CC0, and the free game audio bundles from Sonniss for effects. We dropped a few popular "free" libraries because their free tiers need credit lines or come with conditions.
Then we went one step further: Claude now composes music in code with Audio as Code, written to match the beat map, so the drop lands exactly when the logo does.
For voiceovers we use Kokoro, a free AI voice that runs on my laptop and can be used commercially. One funny moment: it read "mktbox" as "M K T box" until we spelled it out as "marketing box" in the script.
What went wrong, and the checks we added
Every mistake became a rule.
My first video was far too fast to read. Now the pace gets asked every single time, and scenes with text get extra time.
When I tested real footage, my voice sounded like it came through a well. The audio was getting squashed and cleaned too much. Now it stays as recorded, and Claude compares the final sound with the original clip before handing it over.
Endings felt abrupt. Now there's about a second and a half of breathing room after the last word.
At one point we rendered three formats, then one audio fix meant redoing all three. Now it's one format first, the others only once that one is approved.
And moving a subtitle a few pixels meant a whole new render. So we built an editor inside Claude Code: I can drag subtitles, cards and logos, change any text, decide how many words show per subtitle and nudge the timing, then get a quick preview in about three minutes. The small stuff I do myself now.
Real footage works too
Pure animation was the start. Then I wanted to edit videos of real people.
I tested it on myself first: a phone video of me introducing myself. Claude transcribes every word with its exact timing (with whisper.cpp, a free tool that runs on my laptop), cuts the pauses and filler words, checks that no word got clipped, and adds word-by-word captions, zooms and animations that appear on the exact word they illustrate. It can even cut me out of the background so a title sits behind my head.
Once that worked, we made a set of ads with real footage, edited completely by Claude.
Two skills keep it consistent
Everything above lives in two Claude skills: one for animated videos and one for real footage. A skill is a set of instructions Claude follows every time. Mine hold the process, the brand rules, the checks and every lesson from the mistakes above.
That's what makes the tenth video better than the first instead of a fresh gamble. I use the same setup for product videos, launch videos, tutorials, ads and looping demos for landing pages.
What it costs in tokens
I measured this from my sessions. A typical video, from the first message to the final file, uses around 300,000 to 400,000 new tokens. Counting everything Claude re-reads on each turn, mostly the same project files, it's about 15 to 20 million, and that re-reading is cached and cheap. A bigger one with an AI voiceover and music composed from scratch came to about 50 million in total.
How to start
You need Claude Code, which runs inside the Claude desktop app, and a folder for your videos. It's the same setup I used to rebuild this website. Claude installs the free tools itself: one to film the page (Playwright), one to stitch the video (FFmpeg), one for transcripts and one for the voice. It asks before installing anything.
Then start small. Something like: "Make me a 15 second looping video showing how my product works, using this screenshot. Ask me questions first." Answer the questions, approve the beat map, and give honest notes on the draft.
Where this is going
Video and motion graphics used to need a specialist, expensive software and weeks of work. That gap has pretty much disappeared, and I think this is the next massive change for businesses.
The people on X are early. Product Hunt is catching up. The big opportunity is the small and medium businesses that don't use AI yet and have nobody in house who could. They still need videos for their products, their ads and their websites, and now someone who knows this process can make them in an afternoon.