Inspiration

The idea was chosen by the author from nine candidates that another AI (OpenAI Codex) generated from the public hackathon brief.

In my words: Situations like these are close to my own experience, and they do not happen only to me; I often see them around me. At part-time jobs and events, I have had someone take the day off on the day itself and struggled to fill their post. I have seen, or caused, a "domino effect" where moving one person left another post empty. I have looked for someone to cover a shift and passed the request around to many people, and I have rushed to rework duty rosters or shifts in my head or on paper.

From nine candidates that another AI generated, I chose one that would help with a problem at familiar, everyday events. I did not want to stop at one project; I wanted to try by submitting as many as I could.

What it does

Shift Shock is a same-day shift fixer for the volunteer coordinator of a small event with about 6 to 10 helpers.

  • It opens on a "School festival" preset of 7 helpers (made-up names) on a duty board with five posts: Reception, Guide, Stage, Kitchen and First aid, plus a Floating column. Each post needs one person.
  • Click a helper card to mark that helper absent. The post they held turns red.
  • "Find the smallest fix" finds the roster that covers every post and changes the fewest helpers' posts. Moved cards get numbered badges, and each move has a plain-words reason.
  • If a post cannot be covered, an alert panel names the post, gives the reason, and shows the best partial plan.
  • Add helpers, tick skills, set posts. Export the roster as CSV or PNG, or copy it as text.

The preset shows two cases: one absence fixed with 2 moves, and the absence of the only First aid helper reported as "cannot be covered".

How we built it

The author is a Japanese speaker and not a software engineer. In Claude Code, the Devpost Learn Skill Pack's 1-start skill was run in the project folder, and scope.md, prd.md and spec.md were drafted from the Skill Pack's templates. The code was then written by OpenAI Codex, running on the author's own small home computer, from the spec. The app is a static page: HTML, CSS and plain JavaScript, with no build step, no external libraries and no network calls. core.js holds the rules engine: it tries every assignment of present helpers to posts and keeps the one that changes the fewest posts, then uses the fewest Floating helpers, then name order, so the same input always gives the same answer. app.js draws the board and builds the exports. tests.html runs 29 browser assertions (all pass). It is hosted on GitHub Pages and licensed MIT.

Challenges we ran into

Because of the deadline, the author asked the assistant to answer the product questions provisionally; those answers are marked [AI-provisional] in the documents. After submitting, the author reviewed them on 2026-10-27 and kept them all as written.

In my words: The registration and submission fields were in English, and each time I had to choose, I needed a translation and a recommendation. When I was asked about my "confidence" before building anything, I had nothing to judge it by.

I submitted two projects starting about two hours before the deadline, so it was a race against time. I could not make time to answer the planning interview myself, so I had the AI answer provisionally. Right before submitting, saving stopped because of a field's character limit (255 characters), and the steps that needed my own hands, such as switching to the account I use for posting videos, all piled up at the end.

Accomplishments that we're proud of

The app runs fully in the browser with no libraries and was checked in a browser with the included tests.html before the demo was recorded.

In my words: Without writing code, I got both projects submitted before the deadline. Both have a page judges can try without an account, and tests that check how they work. I did not hide the parts I left to AI; I wrote them in the submission text and the planning documents.

What we learned

In my words: If I keep the decisions and the final buttons for myself and split the building work across several AIs, something takes shape even in a short time. I should have finished the steps that need my own hands, such as registration, accounts and publishing, first.

Answering the planning interview at length myself is what makes a project new. Next time I want to start earlier and answer the interview myself.

What's next

  • More than one person per post.
  • Several time slots.
  • Preferences or fairness scoring.
  • Several absences with a suggested priority of posts.

About the "In my words" passages

These passages were added on 2026-11-05, after judging. AI drafted them in Japanese from the project records and common situations, and the author confirmed each line; the English above is a translation. The Japanese originals:

  • Inspiration: 次のような場面は自分の体験に近い。自分だけに発生でなく身の回りで見ることが多い。アルバイトや行事で当日に休む人が出て持ち場の穴埋めに困った。1 人を動かしたら別の持ち場が空く玉突きを見た(または起こした)。シフトの代わりを探して何人にも連絡を回した。当番やシフトの組み替えを頭の中か紙で急いで考えた。別の AI が出した 9 案から身近な催しの困りごとに効く物を選んだ。1 本で終わらせず出せる数を増やして挑戦したかった。
  • Challenges we ran into: 登録や提出の欄が英語で、選ぶたびに訳と推奨が要った。作る前の時点で「自信」を聞かれても、判断の材料が無かった。締切の約 2 時間前から 2 本を出したので、時間との勝負だった。計画の聞き取りに自分で答える時間を取れず、AI に仮に答えさせた。提出の直前に、欄の字数の上限(255 字)で保存が止まった。動画の投稿用のアカウントを切り替えるなど、自分の手が要る所が最後に重なった。
  • Accomplishments that we're proud of: コードを書かずに、2 本を締切の前に出しきった。2 本とも、審査員がアカウントなしで触れる頁と、動きを確かめる試験が付いている。AI に任せた所を隠さず、提出の本文と計画の文書に書いた。
  • What we learned: 決めることと最後のボタンは自分、作る作業は複数の AI に分けると、短い時間でも形になる。登録・アカウント・公開など自分の手が要る段は、先に済ませておくべきだった。計画の聞き取りに自分で長く答えることが、作品の新しさになる。次はもっと早く始めて、聞き取りに自分で答えたい。

Built With

Share this project:

Updates

Submission history