He Had AI Deliberately Write a Bad Story, Then Turn It Into a Movie, and All He Said Was 'Figure It Out Yourself'

Ethan Mollick posted a video on X, with a caption that contained only one instruction to Opus 5.5: "Opus, great, now turn that into a movie, i know you can make it amazing. figure it out." He said that was the entire prompt.

The story starts with the previous tweet. Mollick first gave Opus 5.5 a reverse prompt: "Give me 10 brilliant ideas for a science fiction short short story, pick the most brilliant and execute it terribly." Give me ten brilliant science fiction short-short story ideas, pick the most brilliant one, and then execute it badly.

The story the model delivered was called SIGNAL OF DESTINY: The Message Was A Lawsuit (A Twist Ending). Chad Stellarsen, a scientist at the SETI Institute, and his assistant Brenda receive an alien signal. When they decode it, they find it is a legal complaint document: the Arecibo message that humanity sent to the M13 star cluster in 1974 infringed the other party's trademark rights. The Fermi paradox now has an answer: the universe is quiet because other civilizations chose to stay silent after receiving legal demand letters. At the end, Chad wakes up and realizes it was a dream, and the story closes with the line: "Copyright your stick figures."

Mollick threw this deliberately badly written story back to Opus and had it made into a movie. The finished film was titled CHEAP-O-VISION, with yellow text on a black background, bold with a 3D effect, and a few diagonal beams of light sweeping down from the upper left. The film is 3 minutes 37 seconds long.

Someone asked how many tokens this film used up. Mollick replied that at Opus 5.5's API price, it came to $4.81, but that it accounted for only a small fraction of his Claude subscription.

Another user, AI Adventurer, followed up more specifically: for the 3 minute 37 second finished film, did it come directly from one prompt, or did Opus itself call image, video, and speech models, or use code to draw canvas frames, hook up TTS, and then stitch it together with ffmpeg? How many tool calls were made along the way? Mollick did not elaborate. That gap is more interesting than the answer.

A comment from user Waz was highly upvoted: making something deliberately bad and still funny is harder than making it good. You have to know what good looks like in order to go wrong in all the right places. Another user, Clayton Brown, said that having AI first come up with a brilliant idea and then deliberately execute it badly is the most human prompt, the same approach he used for high school English essays.

Some people compared it with their own experience. User foundrceo said that one sentence was enough to finish a whole movie, while he spent 40 minutes telling the model that the navigation bar still wasn't centered.

In Mollick's tweet, what is more worth looking at than "AI can make movies" is the delegation-style instruction "figure it out." Prompt engineering recedes in this kind of task; what remains is the question of which story is worth filming, and whether you can tell where it is bad in the right way. Making a deliberately bad film for $4.81—film school tuition costs far more than that, and it may not even teach you how.