AI Followed My Plan So Closely It Filmed My Mistakes

0:00 Four still pictures, one prompt, and 30 seconds of video in a single take. A marble pushes the code, the key caps run the tests, a rubber duck reviews it, and a green button ships it. It took three

0:11 tries. On the second try, I told the model not to move my frames. The camera moved anyway. Why it moved is the most useful thing I learned about Cling 4.0. Cling 4.0 makes up to 30 seconds of

0:22 video in one generation. That's twice the 15 seconds of Cling 3.0. On Higsfield, its three modes sit side by side right above the prompt. References takes up to 10 images and five videos,

0:34 so a character, a product, or a look stays the same. Frames takes up to 10 stills, and the model films the motion between them, and edit video changes a clip you already have. Higsfield's own

0:45 comparison reels put Clling 4.0 next to Sea Dance 2.5, drifting cars, glowing creatures, product shots. But a reel shows the best shots. I wanted to test a plan that has to hold together for 30

0:57 seconds. So, I planned a chain reaction on a developer's desk built like a deploy pipeline. Push. A marble rolls down a ruler. Test. It knocks over a row of keycaps. Review. A rubber duck slides

1:09 down a notebook. Deploy. The duck lands on a green button. The screen says deployed and the servers turn green. Every step needs the one before it. If the model loses a single object, the

1:19 chain breaks on camera. I made the four key moments as stills with GPT image 2.5 on Higsfield. Each one framed like a different shot. Then on Higsfield, I open video, pick cling 4.0, and switch

1:33 to the frames tab. The stills go in in order. The first is the start frame, and the last is the end frame. The prompt describes the motion between them. Mine also asks for a camera move. Start

1:43 close, pull back, glide along the desk, then push in on the button. Duration: 30 seconds. Quality 1080p. And before anything runs, the generate button shows the price. 60 credits with 120 crossed

1:56 out. Here are the first two takes side by side. Take one did exactly what my prompt said. Start close, pull back, glide right, push in. But the camera never settles, so the duck, the button,

2:08 and the screen slide into view along the way. And in both, the duck slides off the notebook next to the button, then hops up on top of it. For take two, I changed one thing. A line at the end of

2:18 the prompt saying, "Don't move the frames." The camera still traveled from the marble all the way to the button. So, I measured where each of my four frames shows up in take 2. The close-up

2:28 at 0 seconds, the wide shot at 10, the duck at 18, and the button at 30. That's when it clicked. In frames mode, every still is also a camera position. Four different shots forced the camera to

2:39 travel, and one extra line couldn't overrule four pictures, plus a prompt that asked for a move. So, I rebuilt the frames from one locked camera. All four are edits of the same wide shot. Only

2:50 the moving parts change. The new prompt opens with a locked off static shot. Says nothing may appear, disappear, or duplicate, and lists the only three things that move. New frames in. Each

3:00 one can also be marked as a start, middle, or end frame. Same settings, same 60 credits. Take three. All 30 seconds. One continuous take and the camera never

3:40 moved. Lay the first frame over the last and only the moving parts change. Here's each planned frame next to the moment the model reached it. 9.6 seconds, 16.9, and 28.8 for the end frame. This time,

3:53 the key caps fall in order as the marble rolls past, and the words on the flags and the screen stay readable. Two things still slip and both are my fault. My last frame left out the review flag so

4:03 the model flicks it off the notebook to match and my notebook ends below the button so the duck can't slide onto it. It lands beside it and hops up. Cling 4.0 followed my plan so closely that it

4:14 filmed my mistakes too. Now the second mode, edit video. I pick take three from my generations and add one reference. A red toy robot I made as a still. the prompt. Replace the duck with the robot

4:26 and keep everything else exactly the same. 30 seconds, 16x9, 1080p, and generate. Here's the original and the edit playing together. source reference result. To check what

5:08 else changed, I compared the two videos pixel by pixel, frame by frame. Bright means changed. On an average frame, about 1 and a half% of the picture changes, and almost nine and 10 of those

5:19 pixels sit on the right around the robot. It even gave the robot a red reflection on the desk. The faint marks along the row are the marble and the key caps redrawn a little differently. The

5:29 lamp and the books show no visible change, and the robot follows the duck's path and timing, hop included. So, is Cling 4.0 on Higsfield worth it? For planned shots, yes. 30 seconds in one

5:42 take is enough for a small story. and frames follows a plan more closely than I expected. That's also the catch. It follows the plan, flaws included. Lock the camera, keep every object in every

5:52 frame, and check the physics before you generate. Edit video swapped the character with one image and one prompt, and left almost everything else alone. All four generations together cost 255

6:03 credits. This video was sponsored by Higsfield. Cling 4.0 is live there now, and the link is in the description. Plan it, then generate it. What would you chain together in 30