
For the second week we made one short AI video a day inside one workspace: box art for games that don't exist, a street interview with people who aren't real, a browser game, a debate between models, an anime short, a song and a reel of failed takes. Each day below gives the prompt, the number of tries and what went wrong, including one bug in our own product.
An AI model drafted this article and the ewpire team reviewed it. Every video is AI-generated and says so on screen or in its label.
What changed from week one
After week one, the talking narrator had drawn the most views by a wide margin. So this week we kept the narrator's idea, swapped the quiet kitchen for a louder format, dropped the format that did worst, and linked two days: a game cover on Monday became a real game on Wednesday.
Monday: box art for games that shouldn't exist
Six covers, one short prompt each, first take on every one. Each cover turns an everyday annoyance into a video game genre: MONDAY as dark fantasy, INBOX ZERO as a pixel roguelike nobody finishes, REPLY ALL as survival horror, THE CANCEL BUTTON as an open world with 47 hours of gameplay, AUTO-RENEW as a cozy farm where invoices grow like pumpkins, and REJECT ALL as a 1990s arcade game.
Example prompt: "Video game box art: INBOX ZERO. A pixel-art roguelike, a tiny hero before an endless tower of envelopes. Tagline: Nobody has ever finished it. No logos or rating badges."
The trick was to put the exact title and tagline in the prompt and to say what must not appear. The model also hid a small red cancel button on the horizon of THE CANCEL BUTTON.
Tuesday: a street interview where nobody is real
Three strangers on a night street answer one question: what did your last AI video cost? "No line items. Just one number and a vibe." "All renewals on the first. Payday never stood a chance." "Credits expiring unused. Honestly, the purest kind of freedom." Then the host asks the first one if she is real.
The lines came from an AI chat: we asked for three options per line and picked one of each. Every stranger started as a still portrait, and a video model gave them voices, lip movement and street sound in one pass, four shots in all. A label saying the people are AI-generated runs for the whole video.
The failed take: the video model refused one portrait because the face looked too much like a public figure. We drew a different face and ran it again.
Wednesday: a game made of cookie banners
On Monday REJECT ALL was a cover. On Wednesday Code built it as a game from one prompt, and the first version worked: every Reject brings another banner with a new twist, and Accept stays big and easy.
The game puts ten banners on a made-up local newspaper: a Reject button that dodges the cursor, a preferences screen where one switch turns itself back on before you still have to object to 318 partners, a progress bar that prepares your rejection while Accept stays instant, a notice with the button at the end, and a box where you type I REJECT. With a mouse you need at least 14 clicks. A script played it to the end, so every move of the cursor in the video is real play.
Thursday: three models argue about tipping
Question: should you tip on a takeout order? Three models answered on their own, critiqued each other without knowing whose answer was whose, and one wrote the synthesis. They agree takeout is less of an obligation than table service, and that large or complicated orders give a stronger reason to tip. They split on the plain pickup: no tip, a tip for extra effort, or about 10% anyway. The agreement score was 69 out of 100. It measures how much the models agreed with each other, not whether they are right.
The failed take was ours: the first run showed the synthesis as raw data instead of text. We reran the question and started the fix the same evening.
Friday: an anime short for Friday the 13th
The girl from our home page sits on a rooftop at sunset. A black cat walks along the ledge, and she breaks her melon bun in half and shares it. Seven stills, each animated into four seconds, first take on all fourteen.
One shot jumped from sunset to night halfway through, so we moved it to the end of the cut. Her jacket had a black cat patch before the cat showed up.
Saturday: a breakup song sung by a price-increase email
Lyrics: "Dear valued customer / We're making some changes / Your plan costs more now / Thank you for staying." Two takes of the song came back, and we used the first. The clip has seven stills of the same singer as week one's song and six animated shots, all first takes. One more still came back with fake lettering on neon signs, so we left it out. The lyrics on screen follow the sung timing line by line.
Sunday: the takes that didn't work
A 30-second reel shows the model misses from week one. Asked for 2087, a model drew a lighthouse from any century; asked for claymation, it rendered clean 3D. In our favourite, we asked for a monster that turns into birds, and the hero walked out of the frame with them.
Alongside it, a nine-minute video walks through week one, every prompt and every take, narrated by an AI voice. The voice skipped one line and misread one word, so we checked every section against a transcript and regenerated two.
What we learned
- A format beats a face. Last week one narrator carried the video. This week a street interview did the work, with new faces in every shot.
- Say what must not appear. "No logos or rating badges" and "no readable signs" kept every frame clear of other people's brands.
- Models refuse what they should. A face that looks famous gets refused. Draw a different one.
- Run it twice when it looks wrong. One rerun told a model's answer apart from a display bug in our own product.
- Show the misses. A demo of only the good takes sets the wrong expectation.
If you try one of these prompts yourself, expect a different result: the same prompt rarely gives the same take twice.
Made with AI image, video, music, text and code models on ewpire. Pay as you go. No subscription.
FAQ
How many tries did each video take? Most shots were first takes. The exceptions: one portrait redrawn after the video model refused it, one Consensus question rerun after a display bug, and two narration sections regenerated after a transcript check.
Are the people in the street interview real? No. Every face, voice and street in it is AI-generated, and a label says so for the whole video.
Can I play the cookie-banner game? Not for now. It appears only in the video: it was built in Code from one prompt and played to the end by a script.
Why use a different model for each job? Because each model is good at something different. This week the box art, the talking video, the anime, the song and the game each came from a different model, and none of them would have done all five well.