Twelve AI clay films in five days, for $184
A new channel, 18,971 views, and a $184 generation bill. Here's the pipeline, the ledger, and what twelve tiny films taught us about working with AI.
What we learned
- $184
in cash for all twelve films. Generation was the cheap part: at API prices, the coding agent would have cost more than the video model.
The ledger - 70–76%
of Instagram viewers were still watching at three seconds when the first line named one strange, physical thing. The three office comedies kept 49–53%.
All twelve first lines - 0–4
YouTube views, as of 2 October, for every film posted after the afternoon of 30 September. Five of the first seven had 740 to 1,398. I can't explain it.
The YouTube mystery - Within the hour
an agent usually turned each problem I found into a check the next film had to pass. Every call of taste still came from a person.
Where the agents helped most
Who we are: Sweat AI is an AI-native BPO for banks and fintechs; our analysts review businesses for KYB and work fraud queues. The films were made with the coding agents we already pay for all our engineering.
These are views counted by the platforms; one person can account for several. We collect the counts about every half hour and plot them at three-hour intervals. Hover or drag across the chart to inspect a moment.
Last week, we started a short-video channel with no followers and no ad budget. The cast looked handmade. The crew was mostly code. Claude Code ran the production pipeline, from writing the script to uploading the finished film. I reviewed the cuts; Codex later helped shorten two endings. We made twelve films and posted two or three a day.
Our fal.ai bill for the week came to $184, including rejected takes and experiments beyond these twelve films. An ElevenLabs startup grant covered narration, music, and the first three films' images and video. The coding agents ran on subscriptions we already had. Five days after the first post, the films had drawn 18,971 views and 25 followers. My time reviewing the cuts is outside that bill. The response was more than I expected.
Here is a tour of our tiny film studio: one film from first sentence to final upload, the receipts, what seemed to work, and one mystery that is still bothering me.
The twelve
Nine dramatized stories of real frauds and impostors, plus three office comedies starring Sam, a tired analyst. Hover to play with sound, or tap on a phone. View counts are totals across platforms. Every post is labelled as AI-made; the real people appear as clay caricatures.
What it cost
A finished film now costs $7 to $11 to generate. The more surprising expense was the conversation around it. We used flat subscriptions, but pricing Claude Code's token usage at API rates would put some films as high as $47, mostly because one long conversation kept re-reading its own history. At those rates, the agent directing the film could cost more than the model animating it.
| Production-week spending | Paid |
|---|---|
| Images and video on fal.ai, including rejected takes and experiments (list price of every logged request that week) | $184 |
| Voice, music, and the first three films' images and video on ElevenLabs (startup grant; $112 at list price) | $0 |
| Claude Code on a Claude Max 20x plan ($200 a month) and Codex on a ChatGPT Pro plan, subscriptions we already pay for all our engineering | no extra |
| Cash paid for generation | $184 |
That works out to about one cent of generation cash per view. At API prices, the agents' tokens would have added about $268; paying list price for the ElevenLabs work would add another $112. Together, that would come to about three cents per view. The fal.ai and ElevenLabs totals cover the whole week, including drafts, the model comparison and films not shown here, so they slightly overstate the cost of these twelve. My time reviewing cuts is not included.
Some of the waste was entirely avoidable. A fifth of the generation budget went on rejected takes. One evening, the agents made three explainer films twice over before I had seen a single frame. We discarded $61 of finished work. The fix is wonderfully unglamorous: have a person approve one frame and one clip before generating a batch. The receipt below shows each film's cost at list prices:
SWEAT AI FILM SUPPLY
Elizabeth Holmes
TOTAL$7.14
A receipt for each film, itemised from our own logs, with rejected takes included.
The first three films used ElevenLabs for video before we moved to fal.ai. Their list-price costs were several times higher, though the grant covered the bill. The coding agents use existing monthly subscriptions, so these films added no extra subscription charge.
Where the views came from
Instagram brought almost two-thirds of the views. On TikTok, every film landed between 30 and 250, despite our changes. Facebook contributed 11 in total.
The clearest pattern was in the opening line. Below, the twelve first sentences are ranked by the share of Instagram viewers still watching after three seconds. The top five each give you one strange, physical thing to picture: a press printing money, a drop of blood, a fake prince's dinner. The three comedies and films that began with a wide shot or a setup sit near the bottom. These films had different stories and posting times; we did not run a controlled test. It is a useful clue for the next script.
Then there is the YouTube mystery. Five of our first seven films received 740 to 1,398 views on Shorts. From the Holmes film, posted on the afternoon of 30 September, onward, each has had between zero and four. YouTube Studio shows no notices or strikes. The videos are public, and our upload process has not changed. Instagram did not drop on the same days. I don't know why. If you have seen a similar drop on a new Shorts channel, I would love to hear what you found.
One film, start to finish
Each film starts with a folder and a film.json file. That file holds the script and the instructions for every shot. Those instructions drive narration and image generation. Each image becomes a video clip; the agents assemble the clips into a cut. Here is the Elizabeth Holmes film, step by step.
- The line
Give the first sentence something to show
The script is also a set of instructions. Each short passage of narration comes with its shots, and each shot has an image prompt and a description of the motion. We aim to open on one strange, concrete act, with its object already on screen. Here, the whole promise fits inside one drop of blood.
{ "id": "s01", "vo": "She raised seven hundred million dollars on one drop of blood.", "shots": [{ "id": "h01", "refs": ["holmes"], "prompt": "Close-up: Elizabeth Holmes in her black turtleneck holds up between two fingers a tiny glossy clear clay vial with one bright red drop of blood…", "motion": "she lifts the tiny vial toward camera and the red drop glints; she does not blink" }] } - The voice
Narrate a beat at a time
We generate the narration one beat at a time, transcribe the audio with Whisper, and compare the transcript with the script. An earlier attempt to narrate the whole script quietly changed a line. Now the words have to match before production moves on.
"She raised seven hundred million dollars on one drop of blood."
3.3 seconds · ElevenLabs v4 · each word checked against Whisper before anything else is made - The cast
Draw each character once
Each character gets a reference sheet. We attach Holmes's sheet to every image request for a shot she appears in, to keep her appearance consistent. Her face and hair stay the same across shots, as does the black turtleneck. A new camera angle should not mean a new actor.
- The frame
One still per shot
Nano Banana, the image model, draws each shot from its prompt and the character's reference sheet, at $0.039 an image. Every prompt begins with the same paragraph: plasticine, fingerprints, felt, warm light and a tilt-shift look. That little paragraph does much of the art direction. It is printed in full below, under Making it look handmade.
- The motion
Five seconds of movement
Each still becomes a five-second clip. This film used Kling 2.5 Turbo Pro. After testing ten video models, we switched to Kling 3 Pro at $0.112 a second; nine of the models are compared on the same keyframe below. The motion prompts describe small, physical actions: she lifts the vial, the drop glints, she doesn't blink.
- The cut
Cut to the voice
The voice sets the rhythm. We time each cut to the narration, down to the frame, and add captions two to four words at a time. The music is composed to the film's exact length. Press Sound on to hear the pieces come together.
- The check
Look at every shot
We review eight frames from every shot in the exact file we plan to upload, plus four frames per second from each generated clip. A script checks narration pace, loudness, the length of the ending and more. The upload is blocked if a film fails a check.
- The post
Post it everywhere
The same finished file goes to Instagram, TikTok, YouTube Shorts and Facebook at a scheduled time, labelled as AI-made. Then we look at the numbers before writing the next film.
Holmes so far
1,607Instagram160TikTok3YouTube Shorts0Facebook
Making it look handmade
Before writing our style paragraph, we studied every video on @clayrified, a clay channel we admire. The details we kept coming back to were the ones that suggest a human hand: fingerprints in matte plasticine, sculpting-tool marks, glossy bead eyes under heavy lids, and clothes made of felt and knitted wool. Warm light and shallow focus make the set feel small enough to pick up. We kept text out of the generated frames because spelling was unreliable.
Photorealistic macro photograph of a handmade stop-motion plasticine claymation miniature. All characters sculpted from matte plasticine with visible fingerprints and sculpting-tool marks, big glossy white bead eyes with heavy sculpted clay eyelids, clothes made of felt and knitted wool, miniature handmade set, warm cinematic lighting, shallow depth of field, tilt-shift miniature look. Everything is clay or felt, no real human skin, no real hands. No text, no letters, no writing anywhere. Vertical 9:16 composition.
Three practical rules helped hold that world together. Attach the character's reference sheet to every shot. For an empty scene, say the scene is empty of any living thing
; no people
alone sometimes invited creatures we had never asked for. And put every character and prop needed for the action into the still. Leave something out, and the video model gets to improvise.
For character acting, the video model made the biggest difference in our trials. We tried ten; below, nine animate the same keyframe, side by side and in sync. Kling 3 Pro was our favourite. Watch the hands and eyes: the small gestures do a surprising amount of the acting. One clip per model is a thin basis for a decision, so we treated the choice as provisional and checked it on whole films.
Where the agents helped most
The agents were good at turning measurable mistakes into checks. When I found a problem, the agent usually added a rule to the pipeline within the hour: sample more frames, cap the ending, reject narration that had been sped up after generation. Each correction gave the next film a check the previous one had lacked.
Every film teaches the next one
Each film changes one thing. After it posts, we read each platform about every half hour, find the second where people stopped watching, and turn that into a rule the next film has to pass before it can post.
- 1PostOne change per film
- 2MeasureEach platform, about every half hour
- 3Find the leakThe second where viewers leave
- 4Write the ruleThe next film must pass it to post
Six things the numbers taught us in three days. The rule cards show They sent Sam a pizza menu for proof of address, film 11: the shaded band is what the rule allows, the dot is where that film landed.
- 28 Sepafter film 2
The opening shot
Eiffel opened on a wide shot, with the hook sentence split over two cuts. Mavrodi opened on a close-up of the act, the hook sentence over one 3-second shot.
TikTok: still watching at 3 secondsMavrodi, close-up56%Eiffel, wide shot41%Instagram: average watchMavrodi, close-up37 sEiffel, wide shot18 sInstagram skip rate: 23% for Mavrodi, 45% for Eiffel. Once the scorecard was live, Eiffel held 49% and Mavrodi 76%.
What changedEvery film opens on one concrete act in close-up, and the hook stays on screen long enough to read.
The rule nowHook on screenpass2.9 sRule: at least 2.5 sTime to recognise the object in the opening shot. Our wide openings lost viewers in the first second.
- 28 Sepsame review
The pitch at the end
The 20-second Sweat pitch at the end lost 60% of the people still watching. About 2% reached the last frame.
What changedThe product ending is capped at 6 seconds, from the first mention of Sweat to the last frame. An earlier film ran 11.4.
The rule nowProduct endingpass5.9 swas 11.4Rule: at most 6 sFrom the first mention of Sweat AI to the last frame. Our longer pitches lost most remaining viewers.
- 28 Sepsame evening
Endings that looked broken
The music was composed to a guessed length and chopped where the film ended, and the end card was squeezed to about 3 seconds. Both read as glitches.
What changedMusic is composed to the film's real length, and the silence after the last word is capped. One film had 4.5 seconds of it.
The rule nowSilence after the last wordpass0.45 swas 4.5Rule: at most 0.75 sDead air at the end reads as a broken video.
- 29 Sepafter five films
The first line
The winners' first lines named one concrete, absurd act: "printed his own money", "double your money. With postage stamps." Madoff's first line was abstract, "clients almost never lost money", and did as badly as Eiffel's wide shot.
TikTok: still watching at about 4 secondsBig Bull58%Ponzi54%Mavrodi53%Eiffel36%Madoff33%All twelve first lines are ranked above.
What changedThe first sentence names one thing someone did, not a summary of who they were.
The rule nowFirst lineruleOne concrete act in the first sentence.
- 29 SepTikTok Studio
The first frame on TikTok
In the feed, the first frame was our logo overlay, a still object and a year as the first caption. It read as the start of an ad. TikTok Studio said most viewers stopped watching at 0:01.
What changedThe brand moved later in the film, and the first caption became the hook phrase instead of the year.
The rule nowFirst frameruleThe brand comes later. The first caption is the hook phrase, not a year.
- 30 Sepstill open
The second line
With the opening fixed, TikTok kept 82 to 92% of viewers at one second, so the leak moved to seconds 3 to 10. A second line that adds a new object or a turn kept far more of them than one that gives a name and background.
TikTok: of those watching at 3 s, still there at 10 s (about)New object or turn70%Name and background40%New object or turn: Gignac, Mehta, Mavrodi. Name and background: Holmes, Madoff.
What changedNot a rule yet. Five films is a pattern, not a proof, so new films are now assigned one kind of second line at random.
The rule nowSecond linetestingRandomized test, decided at 13 films per arm.
Five rules came from making the films, not from the numbers
The narration is generated at this pace. Speeding it up afterwards introduced audible artefacts.
Cuts follow breaks between clauses in the narration, as in the clay channel we studied.
The music drops beneath the voice, staying audible without competing with it.
Loud enough for a phone speaker without clipping.
Eight frames from every shot in the exact file we upload.
The pipeline reads views at 24, 48 and 72 hours, the share still watching after three seconds, average watch time, shares and follows. Each film is judged on each platform against the median of the films before it.
Taste was still our job. Every decision to reject a visual style, change a cast or abandon a dull story came from a person watching. The agents' own reviews did not catch those problems. Claude Code built the pipeline and made all twelve films; Codex later re-cut two to shorten their endings. Reusing footage we already had was the cheapest improvement of the week.
We use the same division of labour in our day job at Sweat AI: automate the checks we can define, and have a person sign off on the judgement. That is how our analysts review a business for KYB. The clay studio gave us a small, unusually entertaining place to practise it.