Sizingdeck is a full-stack app built with Claude Code on Elements.
Rooms joined by link with a name, face-down votes, a reveal with the average and spread, and a CSV of estimated stories.
Run this app on your machine. Use this demo as inspiration, or as the starting point for your own app. Your agent creates a new app with this one as the scaffold, and you prompt it to add whatever features you want.
Sizingdeck needed rooms anyone can join from a link, cards that stay face down until the reveal, a table everyone sees change as people vote and a CSV of the results. Each of those is a part of Elements, so the agent spent its 18 minutes on the game itself.
-
A live table. Rooms, participants, stories and votes are LiveTables, one view per room. When someone joins, votes or the facilitator moves to the next story, every open copy of the room updates.
-
Face-down cards. A card reaches the browser only after the reveal, so everyone sees who has voted while the cards stay face down. The reveal shows the average and the spread.
-
Join with a name. A teammate opens the link, types a name and is signed in with a session, in the room a moment later.
-
Facilitator controls as function calls. Adding stories, voting, revoting and saving an estimate call server functions straight from the page with
@rpc. Each call checks the caller's seat, and another participant can take over as facilitator. -
Estimates as a download. The room exports each estimated story with its estimate, average and vote count as a CSV that opens cleanly in a spreadsheet.
-
Data from SQL files. Migrations define the rooms and seed a demo room with five people and seven stories, four already estimated.
Elements is an integrated full-stack app framework and a project server. The project server checks the agent's work as it builds: it type-checks the server and browser code, analyzes the program, applies migrations and runs the tests, then reports the results to the agent. A security mistake, such as calling a server-only function from the browser, is a compile error. The agent reads the report and fixes its own mistakes before it tells you a feature is done. The same checks run on every feature you ask for next.