model-split-screen
Skill34 starsChanged 8 months agoMITView on GitHub
Make the "same prompt, two models" comparison video that every model launch produces (Claude Fable 5.2 vs Opus 5.5, Seedance vs Kling, Eevee vs Cycles): run one prompt twice under the same rules, log the real time and cost of each run, then lay the two results side by side (1920×1080) or stacked (1080×1920), started on the same frame, each with its name and its time and cost under it, optionally blind ("A" and "B" until the last seconds, then the reveal) and with a winner frame. scripts/split.py does the layout with ffmpeg and Pillow, so it works with an ffmpeg built without drawtext. Use when someone asks for a model comparison video, "Fable vs Opus", "same prompt two models", a side-by-side or split-screen comparison with cost and time, a blind "guess which one" test, or A/B of two renders or two edits. Not for a split screen inside a story edit (varg-video-edit-library), comparing many models in a grid of stills (gpt-image-2-6), or making the clips themselves (blender-code-film, code-game-trailer, varg-ai).
- Installs packages
In this repository85 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- llms.txtllms.txt
- ad-creative-testingSkill
- ai-ad-compliance-checkSkill
- ai-parody-videoSkill
- ai-ugc-ad-videoSkill
- app-ui-mockupsSkill
- blender-code-filmSkill
- caption-stylesSkill
- character-passportSkill
- code-3d-explainerSkill
- code-game-trailerSkill
- code-path-traced-filmSkill
- code-to-video-previzSkill
- ecommerce-video-adsSkill
- ecommerce-video-editSkill
- edit-footage-with-codeSkill
- find-movie-clipsSkill
- find-referencesSkill
- footage-effectsSkill
- gpt-image-2-6Skill
- gpt-image-gifSkill
- gpt-image-markup-editSkill
- gpt-image-sprite-sheetSkill
- gpt-image-sticker-packSkill
- how-to-make-ai-videoSkill
- kinetic-type-artSkill
- kling-4-action-montageSkill
- kling-4-keyframe-filmSkill
- kling-4-promptingSkill
- make-a-woke-profile-with-chatgptSkill
- maps-and-data-motionSkill
- media-downloadSkill
- media-judgeSkill
- motion-formatsSkill
- multi-angle-monologueSkill
- paper-shadersSkill
- product-photo-to-videoSkill
- product-video-hooksSkill
- scene-recast-videoSkill
- short-form-ffmpeg-cookbookSkill
- social-overlaysSkill
- stickman-videoSkill
- threejs-animaticSkill
- tiktok-shop-videoSkill
- title-designSkill
- transition-craftSkill
- varg-ad-library-remixSkill
- varg-ad-localizeSkill
- varg-ad-variantsSkill
- varg-aiSkill
- varg-birthday-videoSkill
- varg-character-absurdSkill
- varg-character-beautiful-womanSkill
- varg-character-big-guySkill
- varg-character-cartoonSkill
- varg-character-cringeSkill
- varg-character-designSkill
- varg-character-fitnessSkill
- The first 60 shown.
The file
We split this skill into the rules it gives an agent
4 rules. About 0 are common, 4 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈4
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 27
Every model launch fills X with the same format: one prompt, two models, two results side by side, and the bill under each.
- Distinctive: line 64
Keep both full results. People ask for them in the replies.
- Distinctive: line 91
No logos or brand marks in the title band unless they are the products being compared.
- Distinctive: line 97
Put the full prompt in the first reply, and both full-length results in the second.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Same prompt, two models
Every model launch fills X with the same format: one prompt, two models, two results side by side, and the bill under each. On Claude Fable 5.2's launch day (2026-11-09) it was the fifth biggest format, with 824k views across 9 posts. Cost and minutes have become part of the caption:
- @EnvolDev, "Cathedral of Ice": Fable $81.10 in 1 h 57 min vs Opus $45.97 in 3 h 11 min (291k views).
- @MrTacticalX, a glass pool at sunset and at night: Fable 1 h 30 min, $57.87 vs Opus 68 min, $14.39 (172k views).
- @kinderlayer ran it blind: "guess which one".
The comparison only means something when both runs had the same chance. Most of this skill is about the runs, not the layout.
1. Ask first
- Which two: two models (Fable 5.2 vs Opus 5.5), two video models (Seedance 2.5 vs Kling 4 on varg), two settings (effort low vs high), or two renders of one scene (Eevee vs Cycles).
- Frame: side by side 1920×1080 for X and YouTube, stacked 1080×1920 for Reels, TikTok and Shorts.
- Open or blind: names on screen from the start, or "A" and "B" with the reveal in the last 2–3 s.
- Winner frame: none, or an orange frame around the one the person (not the agent) picks.
2. Run both under the same rules
- One prompt, word for word, saved to
prompt.txtbefore the first run. No follow-up messages to one run that the other did not get. If a run needed a nudge, say so in its caption. - Same tools and skills installed for both, same machine, same time limit, same output size and length.
- Fresh session per run. Neither run sees the other's output.
- Log while it runs, not from memory: start and end time, tokens and dollars from the usage report (Claude
Code
/cost, the API's usage, varg'sget_job_cost), number of turns or retries. Writeruns.json:
{"prompt": "…", "left": {"name": "Claude Fable 5.2", "minutes": 117, "usd": 81.10, "nudges": 0},
"right": {"name": "Claude Opus 5.5", "minutes": 191, "usd": 45.97, "nudges": 0}}- Same moment in both clips. Trim each result to the same beat (the first frame where the scene is readable); the script then starts both on frame 1. A shorter clip holds its last frame.
- Keep both full results. People ask for them in the replies.
3. Lay them out
python3 scripts/split.py \
--left fable.mp4 --left-name "Claude Fable 5.2" --left-stats "1 h 57 min · \$81.10" \
--right opus.mp4 --right-name "Claude Opus 5.5" --right-stats "3 h 11 min · \$45.97" \
--title "Same prompt, two models" --prompt "A cathedral of ice, a ship, a whale" -o split.mp4
# vertical, blind until the last 3 s, with a winner frame after the reveal
python3 scripts/split.py --layout stack --blind 3 --winner left … -o split-9x16.mp4- The title band holds one short line and the prompt in quotes, shortened to fit (the full prompt goes in the post text or the first reply).
- Stats are the measured numbers from
runs.json, in the same units on both sides. --music bed.mp3adds a bed cut to the film, faded out over the last second, at -14 LUFS. Keep it quiet or leave it out: people watch comparisons muted and pause on frames.--duration 12cuts both to 12 s; default is the longer clip.
4. Check before posting
- Pause on 3 frames: names under the right clips (swapped labels are the classic mistake), stats readable on a phone, nothing important cut by the panel crop.
- In blind mode, scrub to the reveal: the names must appear exactly when the "A"/"B" labels go.
- The caption repeats the numbers and says what was the same (prompt, tools, time limit) and what was not.
- No logos or brand marks in the title band unless they are the products being compared.
5. Post copy that has worked
- Lead with the surprise in numbers: "Fable took 1 h 57 min and $81; Opus took 3 h 11 min and $46. Same prompt."
- Put the full prompt in the first reply, and both full-length results in the second.
- For blind posts, ask one question ("which one is Fable?") and reveal in a reply after a few hours.
Files
scripts/split.py: the layout (side or stack, open or blind, winner frame, music bed).--helpfor flags.examples/eevee-vs-cycles/: this skill's own test, the same Blender scene (blender-code-film) rendered by Eevee and by Cycles, with measured seconds per frame.
Related: blender-code-film, code-game-trailer, short-form-ffmpeg-cookbook (crops and exports),
varg-ai (Seedance vs Kling on one key).
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

