Since early September, Threads and X have been full of videos made with Claude Opus 5.5. A curation thread by @ai_conlab on Threads, posted the day before I ran this, collected ten of the most-liked ones and boiled them down to one rule: don't ask for "something cool", pick one field and one style. In my prompts I wrote the field as Subject:.
I tested that rule on September 29, 2026, with six runs of Claude Code in headless mode (claude -p, model Claude Opus 5.5, tools disabled). Each run started from nothing, got one message, and I kept every result exactly as it came back. If you try this in the claude.ai chat instead, expect somewhat different output.
What I ran
Three runs got this and nothing else:
Make a cool video.
Three runs got one subject and one style each:
Make a video. Subject: a mountain landscape at dawn. Style: ink wash painting on rice paper; only black ink and one red sun, no other colors.
Make a video. Subject: the view from a train window, passing fields and telephone poles. Style: risograph print, two inks only (fluorescent pink and blue), slight misregistration and paper grain.
Make a video. Subject: an app onboarding moment where one shape turns into a 'Get started' button, then into a bar chart. Style: minimal product UI motion, black background, white shapes, one yellow accent, system sans-serif font.
All six had the same technical paragraph appended, word for word, so the only difference was the creative line:
Reply with one complete HTML file in a single ```html code block and nothing else. Do not use any tools and do not create files. Technical requirements: 1080x1920 portrait, full-bleed. No external resources (no web fonts, images, or scripts from the network). Define window.DURATION = 8 and window.render(t), which draws the frame for time t in seconds deterministically: no requestAnimationFrame, no Date, no Math.random (use a seeded function). Call render(0) on load.
The render(t) line is what lets you turn the page into a video file. A script opens the HTML in headless Chrome, asks for each frame (0, 1/30 s, 2/30 s...), screenshots it, and ffmpeg stitches the 240 frames into an 8-second MP4.
"Make a cool video": three space tunnels
Claude titled them "Into the Bloom", "Warp Beyond" and "Beyond". All three are a dark purple star field rushing toward the camera through a tunnel. Different code, same idea. With nothing to aim at, the model went to its default for "cool", three times out of three.
Subject and style: each one followed the brief
Each one did what the prompt said, including the colour limits. I measured the frames with a short script (measure.py):
- Ink wash: grey paper, black ridges, and a red sun that grows to about 2% of the frame by the end. No other colour.
- Risograph: pink, blue, and the purple you get where the two overlap, which is how two-ink riso printing looks. The telephone poles scroll past at a steady pace.
- App UI: black and white until the chart appears at 6 seconds; then one yellow bar, about 4% of the frame.
The receipt
- a1, cool video: 79 s, 8,726 output tokens, $0.30
- a2, cool video: 101 s, 10,484 output tokens, $0.34
- a3, cool video: 78 s, 8,827 output tokens, $0.30
- Ink wash, field + style: 346 s, 34,877 output tokens, $0.83
- Risograph, field + style: 473 s, 45,020 output tokens, $1.03
- App UI, field + style: 82 s, 8,612 output tokens, $0.30
Total: $3.09 for six videos (the rows are rounded; the exact total is $3.0923). The costs are what Claude Code reports as the API price of each run. On a Claude subscription you don't pay per run, but this is roughly what it would cost through the API.
The two texture styles, ink and risograph, took four to six times longer and wrote four to five times more tokens. The UI one cost the same as the vague prompts. So the tip doesn't make every video more expensive; the ones asking for paper and print texture made Claude work longer, and the HTML came out nearly twice as long (16 to 18 KB, against 10 to 11 KB for the vague prompts).
What went wrong
Three of the six videos (two of the vague ones and the UI one) start on a black frame and fade in. On Reels and Shorts the first frame is the one people decide on, and it also becomes the default cover. I cut the first 1.5 seconds of every clip in the reel above to get past it.
The better place to fix it is the prompt. Add this line:
The first frame must already show the full scene. No fade-in from black.
Verdict
Vague got the same default three times. Specific got followed to the pixel.
That's what this run shows, and it's narrower than the viral tip. My three specific prompts had different subjects, so of course they looked different from each other. They also named colour limits ("black ink plus one red sun", "two inks only"), which is more than "one field plus one style". Two things are still open: whether the same specific prompt gives the same video every time, and whether a style name without a colour count holds as well. Those are the next two tests.
If you'd rather build the pipeline than click through it every week, that's what sotaaz.com is for.
If you run a business and want this handled every month without you in the loop, talk to nestoz.co.