Model showdown, three AIs on one brief
A fair test people can see for themselves. One brief, three models, first try each, one renderer.
Before you start: read brand.md and MOTION.md if they exist. If not, run brand-intake first, or ask for hex codes, a font and a logo file. Never call a result on-brand without them. The brand applies to the headline and labels, never to the models' outputs.
Step 1: Ask for four things, then wait
- The brief, word for word. Every model gets the exact same text.
- The three models, named exactly as their makers write them, and how the user reaches each one (their own accounts only).
- The logo files for each model, attached. Never redrawn from memory.
- The headline, two short lines. Example: "Same prompt." / "Three AI models."
Step 2: The fair-test rules
- First try only. No re-rolls, no fixes, no "one more go" for the one you hoped would win.
- The models write HTML, you render it. Add this to the brief so no model gets a better export:
Deliver ONE self-contained index.html, [WIDTH] x [HEIGHT] px. Output the complete HTML only. A shared renderer will execute window.seek(t) and export each model's output at the same settings. Implementation: deterministic window.seek(seconds), wrap modulo [SECONDS], every visual state derived from time. Automatic playback when not in ?render mode. Under ?render freeze until seek. Set window.__ready=true after initialisation. No external requests. Do not add controls, debug text, the model name or screenshots.
- Same renderer, same settings for all three: same size, 30fps, same length, same encoder.
- A failure stays in. If a model returns broken HTML or nothing, its panel says so. Never swap in a later try.
Step 3: Stack them
Newsletter (16:9): three panels side by side, each with its model name above it on a plain strip.
Reel (1080 x 1920): three 1080 x 360 panels stacked at y 540, 900 and 1260, over a dark background. Build a transparent 1080 x 1920 overlay.png in HTML holding the headline and the three labels, then:
ffmpeg -y -i a.mp4 -i b.mp4 -i c.mp4 -loop 1 -i overlay.png -f lavfi -i anullsrc=r=48000:cl=stereo -filter_complex "color=c=0x0B0B0F:s=1080x1920:r=30:d=12[bg];[0:v]fps=30,scale=1080:360:force_original_aspect_ratio=increase,crop=1080:360,setsar=1[a];[1:v]fps=30,scale=1080:360:force_original_aspect_ratio=increase,crop=1080:360,setsar=1[b];[2:v]fps=30,scale=1080:360:force_original_aspect_ratio=increase,crop=1080:360,setsar=1[c];[bg][a]overlay=0:540[t1];[t1][b]overlay=0:900[t2];[t2][c]overlay=0:1260[t3];[t3][3:v]overlay=0:0,scale=out_color_matrix=bt709:out_range=tv,format=yuv420p[out]" -map "[out]" -map 4:a -t 12 -c:v libx264 -crf 18 -pix_fmt yuv420p -color_range tv -colorspace bt709 -color_primaries bt709 -color_trc bt709 -c:a aac -b:a 128k -movflags +faststart showdown-reel.mp4
Change d=12 and -t 12 to your length. The crop keeps the centre, so check one frame per panel and move the crop if the action sits off-centre.
What I learned the hard way
- A logo over footage vanishes on light frames. Every label (logo plus model name) sits on its own dark rounded tile in the top-left of its panel.
- Safe zone. The headline sits between y 360 and 520. No word or logo goes below y 1420 or into the right 120px, where the app draws its buttons and caption.
- Label with the full model name, version included. "Opus 5.5", not "Claude".
- End the caption on a question. Mine asked "Which one would you post?" and gave my own pick.
Step 4: Check and export
Pull one frame from the middle. Can you read every label at phone size? Is each logo the right company? Then run the checks in reel-export before you upload.
