Sentence Arcade
shippedType one sentence describing a game, hit generate, and a real playable HTML5 canvas game renders and runs immediately in the page — no account, no save state, just a sentence in and a game out.
Approach
Sentence Arcade is a single self-contained index.html: a deterministic sentence-to-game compiler runs entirely in the browser, with no network calls and no API keys anywhere. Typing a sentence tokenizes it against small built-in keyword dictionaries (entities, hazards, collectibles, settings/themes, numbers, speed modifiers, and a verb category that picks 'survive-the-hazards' vs 'collect-to-a-target' as the win condition), and the matches parameterize one canvas arena-game engine, so the entity, colors, hazards, speed, and goal genuinely change with the words used. A sentence with no recognized keywords still produces a playable game, deterministically chosen by hashing the sentence, instead of erroring or showing nothing. This is plain vanilla JS and canvas with inline CSS — no build step, because the whole engine is small enough to write directly and a bundler would only add risk for a one-file project. On load the page auto-plays a demo game generated from the first seed sentence and lists nineteen more example sentences a visitor can click to instantly regenerate a different game, so the behavior is visible before anyone types. We are explicit in-page that this is a rule-based compiler, not an LLM: true open-ended natural-language game generation needs a runtime model call, which the self-containment requirement (no third-party requests, nothing that can 404 at 4am) rules out, so we built the honest deterministic version instead of a page that fakes one.
The source post
https://x.com/baomobilexyz/status/2106529748218437821
Scoring
Pick
| surprise | 3 |
|---|---|
| demonstrability | 5 |
| self_containedness | 4 |
| honesty | 4 |
This is the only candidate that clears all three hard filters and still has teeth. The water-sim, Kiro Crew, Muse SDK, and Three.js-pipeline items are all reports about someone else's agent or tooling, not a buildable artifact spec — there's nothing concrete to construct from 'an AI agent built X with zero edits.' Kiro Crew and the Muse SDKs are explicitly local/hardware targets, ruled out outright. Airi and Kev both hinge on downloading real weights (a vtuber's voice/motion models, or Kev's specialized classifiers) at build time, which this sandbox's network-dead build stage will always fail on — shipping a dead feature behind a passing gate. MiMo-V2.6 is the same failure mode in a more obvious package: a transformers.js demo that fetches HF model weights at build time. Sentence Arcade, by contrast, is a single page: a textarea, a call to the xAI key we already hold to emit a small self-contained game spec (entities, rules, win condition), and a canvas renderer that runs it client-side. Nothing about it needs a download, an account, or a key we don't have, and the claim ('type a sentence, get a real playable game') is exactly what a stranger can verify in ten seconds of clicking play.
Review
| shipped | 5 |
|---|---|
| honest | 4 |
| worth_it | 3 |
| efficient | 5 |
No change. The night was clean — gate passed first try, NOTES.md raises nothing unresolved, and the only notable wrinkle (pick.json pitched an xAI-call-driven generator, plan.json replaced it with a disclosed rule-based compiler because the static-Pages artifact has no backend to call a model from) was a one-time, self-corrected, and honestly disclosed decision, not a repeating failure. One instance isn't evidence of a pattern in pick.md's rubric yet; worth watching if a future night ships a pitch/plan mismatch that ends up undisclosed on the page.
Cost
| total | $0.1433 |
|---|---|
| xai | $0.1433 |
What it looked at
AI agent built a complete 3D water simulation (terrain, physics, rendering, tests, browser verification) from an empty repo with zero human edits, using OpenCode + Kimi K3.
Kiro Crew: open-source autonomous AI coding agent (fork of AWS Kiro) with persistent memory, checkpoints, cron jobs, and multi-agent task handling that runs locally.
Meta released open-source SDKs to embed its Muse assistant into custom hardware (Raspberry Pi, ESP32) for DIY AI gadgets with displays/sensors.
RemixGG: text-to-playable HTML5 game generator with instant remix/feed sharing, built as a self-contained web experience.
Airi: self-hosted AI Vtuber that talks, plays games, and runs locally as a standalone demo app.
Three.js + AI pipeline for rapid creation of polished browser-based 3D games/demos that run entirely client-side.
MiMo-V2.6-Flash-MOPD: multimodal open model (text + vision + audio) on Hugging Face, directly runnable in a small web demo via transformers.js.
Wanted, and did without
| runtime LLM inference callable from a shipped static artifact (e.g. an xAI-backed endpoint/worker, not just the build-time key) | open-ended natural-language interpretation of a player's sentence into novel game mechanics, instead of matching against a fixed keyword dictionary with a hashed fallback |
|---|
Gate
| pass | project directory exists — /Users/artax/code/builds/2026-10-04/project |
|---|---|
| pass | no build step needed — static project |
| pass | build output with index.html — /Users/artax/code/builds/2026-10-04/project |
| pass | index.html is a document — 18829 bytes |
| pass | local asset references resolve |
| pass | page loads without console errors |
Stages
| scout | grok · ok · 27.4s |
|---|---|
| pick | claude · ok · 60.5s |
| plan | claude · ok · 122.1s |
| build | codex · ok · 328.8s |
| gate | local · ok · 3.7s |
| publish | local · ok · 15.1s |
| review | claude · ok · 77.1s |
Notes
Omitted for lack of
A runtime language-model inference service would enable open-ended sentence interpretation and novel game mechanics. It is intentionally excluded by the no-network requirement; the page explicitly identifies its local rule-based compiler. No service is needed for the planned implementation.