short-game
Skill150 starsChanged 4 days agoMITView on GitHub
Make a short vertical (9:16) video where two AI models build the same self-playing game in lpm, one above the other, and both games then play live in lpm's browser. It is a short-race whose prompt comes from a ready game catalog (3D endless runner, 3D racer, Flappy Bird, Snake, space shooter, Breakout, a GTA-style police chase), with the hook, headline, title and caption written around the game and the brand pair people search for (\"Claude vs Gemini: build a 3D game 🎮\"). Models in, 1080x1920 MP4 + cover + post caption out. Use when the user asks for a TikTok, Reel or Short where models build a game, such as \"opus 5.5 high vs gemini 3.1 pro build a 3D game\", \"Claude vs GPT 6: Opus 5.5 high vs Astra high build a 3D game\" or \"claude vs grok make flappy bird\". A race on any other prompt is short-race.
In this repository22 other files it gives its agents.
- AGENTS.mdAGENTS.md
- mobile/AGENTS.mdAGENTS.md
- website/AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- mobile/CLAUDE.mdCLAUDE.md
- website/CLAUDE.mdCLAUDE.md
- android-cliSkill
- app-review-videoSkill
- app-store-releaseSkill
- app-store-screenshotsSkill
- edge-to-edgeSkill
- lesson-publishSkill
- lessonSkill
- navigation-3Skill
- short-dailySkill
- short-publishSkill
- short-raceSkill
- shortSkill
- lpm-cliSkill
- lpm-configSkill
- lpm-memorySkill
- lpmSkill
The file
Files
A short-race with a game for its prompt. scripts/new.js writes the lesson through short-race's own new.js, so everything in ../short-race/SKILL.md and ../short/SKILL.md still holds (models and CLIs, Run in duplicates, the take, the review) and is not repeated here. This file covers only what a game race adds.
Why games
Creator Search Insights (global, last 7 days, 2026-10-07): "Using AI to build a 3D game" 314K, "making a game with ai" 230K, "ai game generator" 274K, "how to make game with ai" 85K and "ai games to play" 93K (both up 1000%+). For the brand pairs: "claude vs gemini" 267K, "claude ai vs chatgpt" 252K, "claude vs grok" 204K. Races get far more views than feature shorts, and a game that plays itself keeps the reveal moving. The first game race was Opus 5.5 high (12:16) vs Gemini 3.1 Pro (2:24) on the 3D runner.
The games
| key | builds | headline |
runner (default) | a 3D endless runner: lanes, jumps, coins | build a 3D game 🎮 |
racer | a 3D highway racer weaving through traffic | build a 3D racing game 🏎️ |
flappy | Flappy Bird through scrolling pipes | build Flappy Bird 🐦 |
snake | Snake on a 20x20 board | build Snake 🐍 |
shooter | a space shooter blasting enemy waves | build a space shooter 🚀 |
breakout | Breakout with fresh brick walls | build Breakout 🧱 |
gta | a top-down GTA-style police chase in a Vice City-style town | build GTA 6 🚓 |
node scripts/new.js --list prints them. Every prompt in scripts/games.js keeps short-race's rules: index.html at the project root, one file, no libraries or images, a stage with sizes given as shares of it, scaled to fit and never cropped, and "Don't run or test it". On top of that the game plays itself and restarts at once after a crash, so the reveal is live play with nobody at the controls. The stage is 900x700, the shape of a row in the tall window (see One above the other). Races posted before 2026-10-09 ran in columns on a 420x740 stage, and those before 2026-10-10 in short-race's 800x990 window on a wide 900x360 stage: rewrite such a lesson with --force before retaking it. A layout gives each size as a share of the stage (Snake's board spans the height, the shooter's enemies come in from the right). Add a new game to games.js in the same shape (build, play, layout, restart, then the copy fields), not as a one-off --prompt. Keep its whole prompt under about 850 characters: a Claude or Codex copy gets it on a launch line that macOS cuts off at 1024 bytes (see short-race's Traps), and new.js refuses a longer one.
One above the other
short-race runs every race in rows: Run in duplicates with Open side by side split into Rows, run #1 on top and the copy below, from the start of the race to the reveal. Model A's cyan bar sits over the top row and model B's pink one over the bottom, with the clocks and then the finish times, and each row opens its own index.html in lpm's browser.
As in every race, the games play with nothing written over them once the slammed names and the headline go at 1.8 s, and lpm's terminals get one captioned line of about 3 s (short-race's What the video shows).
short-race records every race, real or dry, in a narrow, tall 570x1000 window with the payoff trimmed to the colour bars (see its Traps), so each row's page is about 570x420 pt, a little wider than the 900x700 stage, and each game takes about half the frame. TikTok's caption and sound text cover the lower part of the bottom game, a trade taken for the size.
Make one
Write the lesson:
node scripts/new.js "opus 5.5 high" "gemini 3.1 pro" node scripts/new.js "opus 5.5 high" "gpt 6 astra high" --pair "Claude vs GPT 6" --title "Claude vs GPT 6: Opus 5.5 high vs Astra high build a 3D game" node scripts/new.js "opus 5.5 high" "grok 4.7 xhigh" --game flappyIt runs short-race's
new.jswith the game's prompt and subject, then rewrites the lesson around the game:- Hook: the brand pair short-race names the way people search for it ("Claude or Gemini?").
--pair "A vs B"makes it "A or B?". - Headline: "
<pair>: build a 3D game 🎮". - Title: "
<pair>:<A>vs<B>build a 3D game".--titleuses the user's exact wording (label words such as "Astra high"). - Caption: "
<pair>for coding: making a 3D game with AI.<A>and<B>got the same prompt, a self-playing endless runner, and built it side by side in lpm's terminal. Which one wins? Comment the two models you want me to race next." compare.json:gameandtimeoutMin: 40, or the game's owntimeoutMin(GTA: 60, since Opus 5.5 high took 37:25 on it).
The slug is
<A>-vs-<B>-<game>.--slugand--forcework as in short-race.--game-file <game.json>races a game that isn't ingames.js: the same fields as an entry there plus itskey, or its wholepromptin place of build, play, layout and restart. short-daily'sformats.js newwrites one for the formats it invents. A game meant to stay goes intogames.js.Effort. A Claude side with no effort is raced at high, and the script says so. Otherwise it would run at the user's saved default: Opus 5.5 at xhigh on the runner thought for 21 minutes, hit Claude Code's output cap, resumed, and still had no page at the 40-minute limit. When the user asks for xhigh or max, run it, but warn them it may time out. Gemini 3.1 Pro has one fixed level and takes no effort. A Muse model runs in Meta's Muse Code at its default (high) when none is given, and like a Cursor model it is always model B: "Muse Wonder vs GPT-6 Astra" is written as
new.js "gpt 6 astra" "muse wonder". A model name in--pairthat the voice transcript mishears fails the dropped-words check: see short-race's Traps.Ultracode.
node scripts/new.js "opus 5.5 medium ultracode" "opus 5.5 xhigh" --game gtaraces Claude Code's dynamic workflows (short-race's Models). The ultracode side goes on the left.new.jswrites a same-model hook ("Medium ultracode or extra high effort?"): reword the hook, headline, slam and run line around "Ultracode or extra high?", and settimeoutMinfor the slower side (120 for xhigh on GTA).- Hook: the brand pair short-race names the way people search for it ("Claude or Gemini?").
- Tell the user a take is starting and to leave the keyboard and mouse alone.
Dry run on a trivial prompt. It takes about 2 minutes instead of a whole race:
node scripts/new.js "<A>" "<B>" --dry <scratchpad>/race-dry LPM_TIKTOK_DIR=<scratchpad>/race-dry <scratchpad>/race-dry/dry/take.sh --no-audio --framesIt uses the same models, flow and tall window, a red-square prompt and a 10-minute limit. Check
frames/as short-race says:02-run.jpgshows both banners with the right models, one row above the other, both prompts sent and no intro screen, and05-payoff.jpgshows both pages open with nothing of lpm but the colour bars, the cyan bar (model A) on top and the pink one below, and no project header, address bar or git bar.- The video:
<slug>/take.sh --frames, run with Bashrun_in_backgroundand waited on with an until-loop over its log. A take lasts as long as the slower side: Gemini 3.1 Pro finished the runner in 2–3 minutes, Opus 5.5 high in 12. When a side times out (✗ no page), retake at a lower effort under a new slug. Keep the timed-out folder, and don't post it. - Review as
shortsays, and look at every second yourself:../short/scripts/scan.sh <slug> <out dir>(the out dir must exist). The QA screen check only reads names in Latin script, so a macOS notification carrying the account name in another script (an AirPods battery alert did) passes it.
Finish the reply with each model's time from result.json, then the rendered MP4's full absolute path. To post it, use short-publish.
A popup in the take
A popup over static UI (a pane header, the window's top bar) can be covered instead of retaken:
node scripts/patch.js where <slug> 11.5 # MP4 seconds → record.mkv seconds
node scripts/patch.js peek <slug> 111 115 900,0,900,220 # a tile every 0.5 s → <slug>/_patch/peek.png
node scripts/patch.js apply <slug> 111.9-114.2 1244,0,520,140 111 # span, box, clean frame
node scripts/patch.js peek <slug> 111 115 900,0,900,220 # check: no popup, no seam
<slug>/take.sh --mux-onlyTimes and boxes are record.mkv's own, in pixels of the 2x window capture. The clean frame must come before the popup slides in and show the same UI as the span it covers. apply keeps the untouched take as record.popup.mkv and always starts from it, so a rerun replaces the earlier patches: give every popup in one call. It lists the spans in PATCHED.json. Retake instead when the popup covers anything that moves: a terminal, a game page or the pointer.
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.

