Skip to content

Article ·

Seven AI experiments in one week: every prompt and every failed take

One short AI video a day for a week, all in one workspace: pictures, a fake company, an ad, a song, an anime opening, a game and a model debate. Every prompt and every failed take.

By

ewpire visual: seven frames from a week of AI experiments, from a bakery at dawn to a plane seat. Text: "Seven AI experiments, one week".

We spent a week making one short video a day with AI, all inside one workspace: a picture series, a fake company, a talking-head ad, a song, an anime opening, a browser game and a debate between models. This is the log, including the takes that didn't work.

This article was drafted with an AI model and edited by the ewpire team. The videos are AI-generated, and each one says so on screen or in its label.

What we were testing

Most people pick one AI tool and stay there, because switching costs a new tab, a new account and a new monthly plan. Our question was simpler: what happens when you stop choosing and use whichever model fits the job, one after another, on one balance?

Each experiment had one rule: the prompt goes in the post, and so does the count of tries.

Monday: one word changes the picture

Prompt: "A corner bakery at dawn, 1987, film." We changed one word per frame (the year, the hour, the noun) and kept the rest. Twelve frames, eleven swaps, first take on every frame.

What it shows: when an image is off, swapping a single word moves the picture more than rewriting the prompt.

Tuesday: a fake company in one project

A logo, a one-page site, a jingle and a 10-second ad for the Lost Sock Detective Agency. The logo, jingle and ad were first takes. The site took five rounds of edits.

Six more attempts never reached the site: they ran into two bugs in our own product. One sent a code edit to the image generator when the request mentioned a picture. The other threw away a finished file when the model's answer ended right after it. We fixed both the same day.

Wednesday: an AI narrator against simplicity

An AI chat wrote a 30-second script in the "reasons you don't need..." format; we changed one phrase. We generated two narrators, picked one, and animated three shots. The narrator says she isn't real in her first line, and a label says so for the whole video.

Thursday: a breakup song for a forgotten plan

Four lines of lyrics went into a music model: "You charged me on the first / I haven't called you since the third / Baby, I unsubscribe." Two takes came back; we used the first. The clip took seven stills and six animated shots.

The failed take: our first batch of stills used the wrong reference image, because two files had the same name when we uploaded them. We regenerated the batch.

Friday: an anime opening for a show that doesn't exist

A tired office worker draws a sword made of keyboard keys and fights a monster made of unread email. Every shot started as a still; video models animated them. One still came back with lettering in it, so we left it out. The track had no quiet intro, so the "loading" moment before the drop is an edit, not the music.

Saturday: a game from one prompt

Prompt: "Build a browser game: try to cancel a subscription. The Cancel button runs away." The first run returned a working game: a made-up cookie club, a button that dodges the cursor fourteen times before it gets tired, seven "Are you sure?" screens and a retention bot. It took us 27 attempts to escape.

Sunday: three models argue about a plane seat

We asked three models: is it rude to recline your seat on a plane? Each answered alone, critiqued the others without knowing whose answer it was, and one synthesis came back. They agree reclining isn't rude in itself. They split on whether to ask first. The agreement score was 66 out of 100. It measures how much the models agreed, not whether the answer is right.

What we learned

  • Frame first, then motion. Every animated shot this week started as a still. A still costs a fraction of a video, so the cheap step absorbs the failures.
  • Count the tries out loud. "First take" and "five rounds" tell the viewer what to expect from the same prompt.
  • Different jobs, different models. The music, the stills, the animation and the code each came from a different model. None of them would have done all four well.
  • Making things finds bugs. Two of this week's failures were ours, and we only found them by using the product the way a customer would.

All seven videos are on our channels, with the prompts in the posts. 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, music, video, text and code models on ewpire. Pay as you go. No subscription.

← All articles