100% of this document was written by Po-Shen Loh, without any AI (even for
editing/feedback). It was then fed into Claude Code to build the initial
page.

PLAN:

This page contains a blog post from me (Po-Shen Loh). It should go into
poshenloh.com/posts/20260904-ai-video-ad-campaign. Also make a version at
poshenloh.com/posts/20260904-ai-video-ad-campaign-zh which is translated
into Chinese at a high quality translation level. If you auto-detect the
browser language is Chinese on the original main page, auto-redirect to the
Chinese version. There is no need for a button to change the language to
Chinese on the main page. Automatic is enough.

The page should be more fun than my usual page, because it corresponds to
ads which are seriously fun. Look in
~/ads-live-motion-assets/out/spinoff-c2.mp4 for an example, and also look
through the other spinoff*.mp4's too, which are in various stages of
polish. And look in ~/ads-live-motion/ to see how they were created, to get
their visual language. I wonder if any of that visual language can be used
for this page too. The colorful rectangle-shadows behind the embedded
videos might be too loud for this page, but see if you can come up with
something that jives with the aesthetic. For example, the brain bumping the
phone out of Thought Full. Make sure to use the same visuals as the ends of
the ad videos we have. That should be easy because those were code anyway.
But don't include the "Math made" there, or the link
poshenloh.com/thought-full. Just the "thought full" is enough with Mission
and Newsletter.

This is a very long post. Some parts are prompts I made to Claude/Codex.
Think about how to display everything in a more delightful way. I
have also interspersed philosophy in the post. I originally was thinking
that there could be a way to have the philosophy be in smaller font and 50%
bright, where if you click on it, it expands to normal size text and 80%
bright. But I also don't know if there's a clean way to separate out
philosophy in the way I've written things before.

I do think there should be a TL;DR summary somewhere which identifies the
key steps that I took which unlocked the possibilities, and lets you jump
to the part of the page which talks about them (smooth-scroll quickly down,
with easing in and out). I suppose that's like a navigation bar of sorts,
but I already have a navbar for my personal website. Some of those key
things might be:

* Remotion
* Lyria
* Storyblocks
* Teach Claude what's so special about us to give itself full confidence
* Blur text using baseline detection

Number them without leading 0's.

Don't include any links back to our LIVE platform or courses at the bottom
of this page. But do include a link to the mission, and to subscribe to my
newsletter. Look at the contact page in the poshenloh-www repo for that. If
the footer already shows by default, then no need for links to follow my
socials.

UI: make the video ads play/pause when you click on them, not only on the
tiny play button. On desktop, when you click on a video, make it pop up to
take up a good chunk of the screen, with an X button on the top right that
appears when the mouse moves over the video block. I hope the HLS is smart
enough to figure out which resolution to send, so that desktop people get
a high-res version.

Credits: Include all significant tools, but not things like the font.

Fine-tuning feedback:
* In the subtitle, don't switch between italics and not-italics.
* Whenever you use the typing-in animation (e.g., like "in a Weekend With
  $100"), don't use a vertical cursor bar, because it messes up centering,
  and it also sometimes leaves artifacts when it stays blinking in the
  wrong place. And while it's typing, make sure there's no jerk. There was
  jerk when the first W was typed.
* In the video player, make sure the buffering gets ahead of the video
  before first playing.

***

TITLE: How Math + AI Made an Ad Campaign in a Weekend With $100

DATE: Sep 4, 2026

[Note: I actually wrote this text myself (link to this document placed
somewhere reasonable in public/), and then asked Claude Code to build a
webpage around it. The initial prompt, including the original post text
itself, is also in that file.]

The goal of this post is not to explain how to prompt AI, but rather how to
think, problem-solve, and use AI responsibly to bring a complex human
vision to life, without blurring reality and fiction.

Last week, I needed to design a campaign with a whole series of video ads,
for advanced math classes that are free for kids in schools that are rural
[link to
https://okcfox.com/news/local/new-live-program-brings-after-school-math-and-science-coaching-to-oklahoma-schools]
or 85%+ free/reduced lunch [link to
https://www.santafenewmexican.com/news/local_news/santa-fe-students-offered-unique-opportunity-through-global-math-program/article_73f44048-e277-11ef-bd9c-7b182ed187d9.html],
and reasonably priced for everyone else. My vision for the campaign was:
what if Apple, Buc-ee's, and Michael Jackson had a collaboration? There was
only one problem. I had no designer and no budget.

I did have an example that I liked, custom-designed over a year ago by the
talented Julia Du. She worked with us part-time between her graduation from
Parsons and starting work at Google.

[Don't link to Julia's video here, because we should lead with ours.]

Thanks to Claude, Codex, Remotion, Storyblocks, Google Lyria, and
Elevenlabs, we have
an entire ad campaign. Of course, it's not as good as what a professional
designer can create. But it seems to already outclass every other ad
campaign in the after-school math enrichment industry. And those companies
actually have enormous marketing budgets, while we give away our profits to
teach kids in lower-income and rural areas.

[Show all of the portrait ads here in a horizontal carousel:
video-hosting-pipeline:www:20260904-ads:{c2,e,f,d,g,p,n,m,o,q}.mp4, where
c2 is centered and first. So presumably there is nothing to the left of it.
]

Since I am an educator, I thought I'd share the recipe of how this was
done. Note that this task is different from video editing (which many
report you can ask GPT-6 Astra to "just do" for you using your regular
video editor). This is more akin to composing a symphony like Beethoven,
than mixing tracks like an elite audio engineer.

I knew that if I asked AI to "make a video ad", I'd likely get something
generic and boring. (There is a mathematical reason which supports this
intuition, which I'll explain below.) I didn't bother at the time, but just
for fun, I tried that just now, with this prompt:

> I'm curious what you can create. Make a video ad inspired by this one.
> Here are our unique value props. It should be AWESOME.

[I uploaded Julia's video (link to
video-hosting-pipeline:www:20260904-ads:202503-julia.mp4) and the 5-page
document of our unique value props, but when clicking on the button to view
my document, don't scroll to the bottom of the page, because people will
have trouble scrolling back up. Maybe make it a pop up?]

It made this:

[video-hosting-pipeline:www:20260904-ads:202503-julia.mp4 next to
video-hosting-pipeline:www:20260904-ads:gpt56-high-ad.mp4, captioned
appropriately]

Oof. It feels generic (much more on that later). The whole point of the
Julia's video was to make the text playfully appear across each frame. That
was entirely lost. And ChatGPT had no way to get videos of real people,
which lost the human feel.  At least it didn't AI-generate fake people. The
visual layout and polish look sloppy. The words themselves might not be
terrible.

In case you're curious, several days later, I also tried using OpenAI Codex
with GPT-5.6 Sol High Effort, as another test after I had succeeded already
with Claude. Even with prompting, and will full access to Claude's codebase
and our later-generated audio tracks, it produced these two iterations.
It's always valuable to explore and find out if there are better ways of
doing things, but after the experiment I switched back over to Claude
Fable 5.1.

[video-hosting-pipeline:www:20260904-ads:codex-ad1.mp4]
[video-hosting-pipeline:www:20260904-ads:codex-ad2.mp4]

There is a mathematical reason why the task of AI-creating videos is hard:
the number of dimensions in the space of videos is absolutely enormous. At
its most basic, if you simply consider a 30-second 1080x1080 60fps video,
that is 30×1080×1080×60 = 2 billion pixels, each of which has a red, green,
and blue value, which is a total of 6 billion dimensions. If each value
ranges from 0 to 255, that's 256^{6 billion} possibilities to optimize
among.

In contrast, coming up with an English language 150-character tweet "only"
has around 30^150 possibilities. The exponent is the killer. In reality,
the objective of AI algorithms (and perhaps our own human intelligence) is
to vastly reduce the dimensionality in both situations. AI companies almost
definitely have their models focusing on reducing this dimensionality. But
there is such a richness of ways to direct the reduction process. How would
the AI model know what you really have in mind, to steer the focus?

Perhaps that's why AI-generated video often feels "generic": it takes
impossibly long to search through and compare all of the possibilities. Yet
what sets any product apart is to be iconic. I once heard a marketing
lecture where the speaker said you shouldn't ever try to tell people you're
the "best at X". You should be the "only one who does X." And that requires
insight.

In this case, the insight was to reduce the dimensionality by asking Claude
to find a programming language which could fully specify the whole
animation in words and commands. Note that I did not first ask AI to create
the ad. I asked to research the ideal way to vastly compress the number of
dimensions.

> Open ~/ad.mp4 and look at its style. Then look at the files in ~/julia.
> Those are the ingredient clips. I'd like to generate our own spinoff ads.
> I wonder if it's now possible to create them in some web format and
> render them ourselves and create MP4's instead of using Adobe After
> Effects. Because then we could specify everything programmatically. Is
> this actually possible? If so, what would be best? Next.js? React? Vue?

[These are actually my real prompts. Apparently I still capitalize and use
punctuation. I think it is respectful to do that. Who knows, maybe if
you're respectful to AI, it might try harder for you too.]

The answer was terrific:

=== FROM CLAUDE COWORK ===

**Answer: yes, and the framework question was the wrong axis.** Next.js is
a web-app framework with no notion of a frame. What the job needs is a
*deterministic frame renderer* — where the picture at t=3.2s is a pure
function of frame 96, so headless Chrome can be stepped frame by frame and
piped into FFmpeg.

Researched the field and built two working spinoff ads as proof rather than
asserting feasibility. Options considered:

| option | verdict |
|---|---|
| **Remotion** | chosen — shipping, mature, best-in-class audio sync |
| HyperFrames (HeyGen, Apache-2.0) | real contender, no licence fee, but 0.7.x and fiddly audio |
| Revideo (MIT) | 0.x; OSS repo trails a commercial product by the vendor's own admission |
| Motion Canvas | out — no stable release in ~20 months, no headless render path |
| Editly / fluent-ffmpeg | out — fluent-ffmpeg archived May 2025 and self-declared broken |
| DIY Puppeteer capture | out — means reimplementing deterministic virtual-time capture |
| Shotstack / Creatomate / JSON2Video | out — hosted-only, rent-per-minute at variant scale |

=== END CLAUDE COWORK ===

And, the first result it produced was already surprisingly not awful!

[video-hosting-pipeline:www:20260904-ads:a.mp4]

Particularly because I was not evaluating it based upon its final quality,
but rather I was tracking progress along the pathway of reducing
dimensionality. I knew that once I could replicate the capabilities of
Adobe After Effects, composed through a Claude Code pipeline in Remotion
[link to them] (the first huge part of solving this problem), I would be
able to generate a vast number of videos. And this particular output showed
that it was possible to generate professional-quality text animations
through code.  Then I asked:

> This is really good. But the things you made weren't as spiffy as what
> she made. Can you attempt to clone hers as well as possible? Also our
> brand font has since changed to Plus Jakarta Sans which is open source

That produced:

[video-hosting-pipeline:www:20260904-ads:live-ad-clone.mp4]

I was thrilled to see that my nascent system could not only bring in all of
the necessary components (animations and video), but it could understand
Julia's video as an example piece, and learn from it. Notably, it got this
far from **one single example**, not a ton of training data. That indicated
to me that I had found the correct language for expressing Julia's video.

I then also knew that this was going to take off, because I could then
leverage AI and computers for what they are good at. We humans are very
slow at clicking. It is a painstaking process to zoom in, frame-by-frame,
and touch up or align images/sound to be just perfect. AI can speed up the
clicking process. The creative process and task of artistic direction still
remains, but it is hyper-accelerated because what was once laborious is now
super fast. I can ask my system to make new version after new version, and
iterate quickly.

So it was time to iterate. For that, I needed to track all of the editions,
because progress would likely be some steps forward and some steps back.

> This is really good. I've created a Git repo called Expi/ads-live-motion.
> Please put the clone source code and all generating stuff into there. Of
> course not the rendered output. And make sure you have instructions and
> some record of my commentary.

At that point, I also realized I did not need anything from MacOS
(initially I wasn't sure), and so I asked if it was time to switch away
from my Mac Mini M1.

> I'm thinking of switching this work to Claude Code on a AMD Ryzen 5 9600X 6-Core
> Processor running Linux. Can I do that, or do I need the full power of Cowork?

Once I switched over, progress sped up dramatically. In part because my
main Linux workstation has 5 monitors, arranged vertically.

[public/images/posts/20260904-ads/5-monitors.jpg]

Next up was sound. In the past, we had just bought a soundtrack and used it
over and over again. Emboldened by what I had gotten in such a short time,
I decided to go for a custom soundtrack. I remembered from conversations
with people in the entertainment industry that a sign of true quality in
movie production was a custom-composed score that complemented and deepened
the emotion of what was on screen.

> OK this is pretty good. Are you able to generate your own soundtrack too?
> If so, please put it on the spinoff-c.

Unfortunately, the soundtrack was laughably bad. It sounded like a 1990's
video game. Despite asking Claude to try harder, it didn't improve much.
That's when I realized that I needed to reduce dimensionality by searching
for a tool that would turn text into a musical production.

> The visual is good. But the audio is not. I think the reason why the
> visual is so strong is because you're using a really good tool for
> rendering the video with very high quality effects. Search for a
> similarly good tool for composing and then synthesizing the audio, which
> can even take into account timings like what you did here. Of course, the
> sound effects would be layered on separately.

It found Google Lyria [link] and ElevenLabs Music [link]. Both were less
than $10, and so I got access to both of them, and passed their API Keys to
Claude Code. It then produced a whole slew of soundtracks for me to listen
to. But that was inconvenient, because I had to imagine how the video
played along with the sound. So, the natural next request was to:

> Can you whip up a simple web app I can use to preview the audio and video
> together?

It took less than a minute, and even had some style.

[public/images/posts/20260904-ads/listening-room.webp]

With this user interface, I then engaged in 7 rounds of providing feedback
to train the music. Music is also extraordinarily-high-dimension. I
realized that the way to reduce that dimension could be to establish 3
phases of moods: starting pensive, then piqued-interest, and concluding
triumphantly. I remembered listening to one of my college roommates from
20+ years ago, who was really into the theory of music, and I remembered
that there existed things called chord progressions, minor and major keys,
and key changes. I didn't know much about any of those things, but that
general knowledge gave me things to prompt about. Finally, I had a general
engine for producing a variety of soundtracks to use, which are those that
appear in the videos I showcased at the beginning. (Those audio tracks are
not final, and I'm still experimenting with instructions, particularly
around chords and keys.) [So when writing the Lyria unlock, don't mention
key changes because that's not done yet.]

From there, I moved back to the visuals. I generally dislike AI-generated
visuals, because I don't want to show fake things. Through discussion with
Claude, I learned that it was able to search stock footage on a platform
called Storyblocks [link]. I then asked it to research all of the possible
providers:

> I would be happy to get released clips from a paid library. This is going
> to be a paid campaign. But the pricing needs to be good.

The two main contenders ended up being Adobe Stock and Storyblocks.
But crucially, Claude was not able to search Adobe Stock, while it was able
to search Storyblocks without login. (Of course, it couldn't get the high
quality video files, but it would be able to sift through footage and
recommend video clips.) That became the deciding factor. Claude Code's
searching ability, combined with its ideation of each ad's story, would
provide the necessary real human footage to complete each piece.

At that point, I had settled into a design framework for each piece:
whenever we talk about our classes, we would use actual screen recordings
from real classes (not staged and definitely not AI-generated)! And
whenever we show pain points in the rest of the world, we would use stock
footage. All of this would be interlinked with playful text animations:

> Also, learn from Julia’s video, which still feels more refined and
> creative in some points. Like she had the idea of panning stock footage.
> And she had many playful effects from Adobe After Effects. We don’t need
> to clone all of them. But just like how I suggested using the copy paste
> animation in C, think about how to have clever animation choices support
> the story. I don’t think we need gratuitous distractions. But we can be
> clearly deeply thoughtful. I want the ads to tickle the intellect and
> make people say, wow, this is so deep. Yet so in tune with modern kids
> and entertainment. Like how a Disney movie has many parallel meanings
> that appeal to kids and adults in different ways. By the way, record
> somehow this philosophy so we can have an iconic ad series. Actually,
> very good if you can help brainstorm what it is that would be iconic.
> Just like how Buc-ee’s has their iconic style. Remember that we are
> aiming for iconic because as you have researched yourself, we are head
> and shoulders above the rest.

Oh, a note about the last sentence in the prompt. My understanding from
reading Claude's Constitution [link to
https://www.anthropic.com/constitution] (yes, I read the whole thing) was
that Claude is not supposed to deceive. So, I was fairly sure that Claude
wouldn't go all-out unless it was itself convinced that there was something
special here. To that end, I had previously passed in a document which I
had written some years ago, about the unique value propositions of our
classes [make it possible to click and open public/live-unique.md on this
page, also with a link to that exact file for SEO], and also specifically
asked Claude to do its own independent research to develop the necessary
conviction:

> OK this is reasonably good. Please sharpen the words. Remember, you are
> making ads for literally the best product in the market. To make you more
> confident of that, please do a deep research yourself into what else
> there is in the market for beyond-curricular math classes that are taught
> online. Scour YouTube too. You are allowed to use Google Chrome. If you
> need me to connect something, let me know what to click.

By the end of this, Claude was spitting out ad after ad after ad. I even
got it to create in Chinese.

[video-hosting-pipeline:www:20260904-ads:c2-zh.mp4]

And since
I had settled on the correct languages to reduce dimensionality for
visuals, storyline, and music, Claude was able to synthesize entire
campaigns automatically, while I could interact with normal conversation
and nudge it along to make definitive improvements. For example, since I
had the correct language, Claude could reliably change the timing, color,
or position of anything I asked in plain English. My goal of Buc-ee's meets
Apple meets Michael Jackson was coming to fruition.

There were a few more things I did. Most notably, since we showcased real
screen recordings of classes, I had to redact students' names. That is as
painful work as removing a mustache from a Hollywood film [link to
https://www.buzzfeed.com/delaneystrunk/henry-cavill-had-his-mustache-digitally-removed-in].
In the past, my team would just put up a big blur rectangle. Now we can do
this:

[video-hosting-pipeline:www:20260904-ads:live-62-redact.mp4]

I originally just asked to blur out the last names. But that was
inconsistent. The final success came from asking OpenAI Codex GPT-5.6 Sol
High Effort to perform a Goal Seek to remove all last names, with this
instruction on how to precisely blur out each name:

> For each last name, identify the text baseline. Blur out the rectangle
> whose bottom is 30% below the text baseline, top is 15% above the top of
> the text line, left is one pixel after the end of the previous word, and
> right is one pixel before the start of the next word (or if there isn't
> one, one pixel after the end of the last name). Do not use this
> computer's OCR system because your visual system is better.

[Note: this is the synthesis of several prompts I asked Codex.]

FIX: I used Codex because it appears to be better at understanding images
than Claude. Definitely GPT 5.6 Sol High was better at that than Opus 5. (I
had run out of Fable 5 credits.) So I also assumed GPT 5.6 Sol High would
be better at making the final video. But it turns out it was much worse
than Claude. Maybe Claude is like Beethoven, who could compose even after
he was deaf.


THOUGHTS

I never thought I'd make a post teaching how to make videos. I majored in
pure math in college, and all of my later degrees are in pure math. But it
seems to be that in today's AI world, the pure math intuition for
complexity and its reduction, when combined with a wide base of general
knowledge/experience, is very useful.

Am I as good as a professional designer? Absolutely not. And I always
admire people who explore the limits of creativity and ability. But my
space of possibilities of what I can do (with very limited budget) has just
expanded dramatically. For example, now I'm considering starting a video
podcast series, because I can build an automatic pipeline for editing and
sharing highlights.

Where, then, is the person, or is that just making more AI slop? Well
actually, in this particular process, the entire goal is to create
something iconic, not generic. It remains extraordinarily valuable to be a
thinker who can break down big things into smaller parts, because that is
what enables you to communicate your vision effectively. Generic things
emerge from the lack of vision. Be iconic. Figure out what really is you.

In some sense, the core of this whole campaign came from our decision years
ago to fuse the worlds of math stars, professional actors, Internet
live-streaming, and social impact. That's what took me on road trips all
across the country, on my speaking and listening tours [link to /wsj-tour],
to help people get ahead of the advance of AI. Those tours led me to
discover the wonders of Buc-ee's, which I took photos of on my iPhone, in
between Michael Jackson songs.