AIToday
Large Language ModelsAI Coding AssistantsZenn AI/MLPublished: Nov 14, 2026, 04:00 JST

Tosh1 ships Hoshizora Dash, rebuilt 68 times in 3 days

Send on LINE
Tosh1 ships Hoshizora Dash, rebuilt 68 times in 3 days

3 Key Points

  1. What happened

    Tosh1 published "Hoshizora Dash", a browser game with 4 modes, 4 enemy types plus a boss, 13 achievements and a ghost-run feature, rebuilding it 68 times (v1 to v68) in about three days.

  2. Why it matters

    The time went into quality control, not code generation, with most bugs caught by playing on a real phone; Tosh1 says an AI-only build would have finished on day one.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Context & Analysis

The production log focuses on what broke rather than a launch announcement. Tosh1 set up a division of labor from the start: the human decided the concept, prepared assets and judged whether the game felt right on a phone, while the AI named Muse did implementation, verification and publishing operations. That arrangement meant most of the bugs were found by playing the game, not by reading code.

The asset pipeline is where the project's hidden cost sat. Tosh1 started from a single sprite sheet with 43 poses, cut each pose out, made backgrounds transparent and fixed a common standard for size, foot position and scaling before adding anything else. That standard made it possible to swap in a second character using the same 43-pose layout. But the game actually required 57 sprites per character, and forgetting to include the extra 14 in the published asset set caused the flicker bug described later.

The failure stories share a theme: verification discipline. A horizontal-mode start bug could not be reproduced in the test environment and only got fixed once the CSS conflict was found on a real device. A staff-roll loop broke because a marker value of −1 was confused with a negative scroll position. Tosh1's conclusion is that "fixed" should only be said after reproducing, correcting and rechecking under the same conditions, and that comparing the list of files the game requests against what is actually shipped should be done mechanically before publishing.

FAQ
What tools did Tosh1 use to build the game?
Muse handled implementation, verification and publishing operations. ChatGPT and Gemini were used for some assets and backgrounds, and the browser game runs on GitHub Pages.
What went wrong after publishing?
The published version showed placeholder characters because sprite folders were missing from the repository, and a later flicker bug traced to 28 missing files (14 frames for each of two characters).
Did it cost anything to make?
Tosh1 says it was not entirely free: on November 9, 2026, about 25 points of a free weekly AI quota were used in one day, and on November 11 usage went from 26% to 72%, with some weeks exceeding the free allowance.
Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleAIエージェント can automate four HR and recruiting tasks, but people must decide