Tokengeoji (토큰거지)

Inspiration

AI requests usually disappear inside private chats, even when the answers could help, entertain, or inspire other people. Tokengeoji turns those requests into public community prompts: one person asks, other people post AI-made results, and everyone can react, compare, and remix what appears.

What it does

Tokengeoji is a Korean-first AI request-and-result community. Members can:

• post a structured request with context about what they need;
• add text, links, images, or videos as public results;
• react to results, accept a helpful result, or remix a request;
• build a profile around the requests and results they contribute;
• use a same-prompt benchmark lane to compare model outputs in a repeatable format.

The first request is frictionless. Repeat posting and editing use a lightweight ad-gate ritual that supports the service without turning the community into a marketplace.

What I extended during Build Week

Tokengeoji existed before the submission period, so I focused the Build Week work on a meaningful product extension. From July 14-15, I:

• reframed the product around clear, everyday AI asks instead of technical AI jargon;
• added an ask-situation picker so people can explain why they need help;
• built a personal same-prompt benchmark flow with reusable comparison briefs;
• improved the public profile, system-host labeling, help, login, and moderation-facing flows;
• hardened the production Docker path and the Korean-first earth-and-hanji workbench interface.

The dated commit history separates this Build Week work from the pre-existing community MVP.

How I built it

The app uses Next.js App Router, TypeScript, Prisma with SQLite, Firebase Authentication, NextAuth sessions, and a server-rendered community feed. Vitest covers domain logic and Playwright covers the core request, result, profile, authentication, and moderation journeys.

Codex helped me inspect the existing product, turn product decisions into scoped changes, keep the UX contract tests aligned with the implementation, and verify the release path across the repository.

Challenges and lessons

The hardest part was making structured AI comparison useful without turning the site into a technical dashboard. The solution was progressive disclosure: everyday language first, with model and benchmark details available only when they help.

I also learned that an ad gate works best close to submission, after the user has already shaped a request, and that public AI results become more valuable when the community can react to and remix them instead of treating them as private deliveries.

Built With

• next.js