Inspiration
Twitch communities are full of terrible ideas—in the best possible way. I wanted to turn that creativity into something a streamer could actually fight.
RAIDFORGE started with a simple concept: chat invents the attacks, Codex turns them into a playable boss, and Twitch raids carry that monster into another community.
Instead of leaving a stream with only a clip, a community could leave behind a monster with its own creator credits, mutations, and history.
Meet The Honkfather: a goose in a fedora with a gold chain and absolutely no intention of playing fair.
What it does
Chat builds the boss. You survive it.
RAIDFORGE combines a boss-fight arena with a workshop for community-created attacks.
The connected workflow lets viewers propose mutations using !forge and vote with !vote. The streamer selects an idea, Codex generates a bounded attack program, and the streamer previews and approves it before equipping it.
The fight then becomes a shared challenge:
- Three escalating phases introduce targeted floor strikes, aimed volleys, and crossing attack lanes.
- Chat powers let viewers vote for a shield, healing, or faster enemy projectiles.
- Overdrive rewards close dodges with a temporary damage boost.
- Shareable challenge links let friends face the same boss and settings.
- Run reports capture personal bests, accuracy, creator credits, and downloadable scorecards.
- Raid handoffs are designed to transfer the boss and its history between participating Twitch channels, with acceptance required before replacing the receiving channel’s boss.
The public demo includes playable local practice, preset mutations, challenge sharing, and run tracking.
How I built it
I built RAIDFORGE with Codex, using JavaScript, HTML Canvas, and CSS for the game and responsive workshop interface. The goose artwork, projectiles, warning zones, and combat effects are drawn directly on the canvas.
A Node.js companion service, built with Express and WebSockets, handles the connected workflow, room state, spectator frames, and Twitch integration code. Runtime attack generation uses the OpenAI Codex SDK.
Generated attacks use a small declarative language rather than executable JavaScript. Each program is validated against instruction and projectile limits, then presented for streamer approval.
The source is hosted on GitHub, and the public frontend is deployed on Vercel. Live Twitch and Codex features require the separately hosted companion service and configured credentials.
Challenges I ran into
The biggest gameplay challenge was making the boss difficult without making it unfair.
Early fights could end before the attack patterns became threatening. I redesigned combat around escalating phases, attacks that punish standing still, and visible warnings that give players time to react. Phase transitions clear existing threats and create a brief opportunity to reposition.
Another challenge was preserving community creativity while keeping generated attacks predictable enough to run safely. A constrained attack language and an explicit preview-and-approve step made that possible.
Mobile play also needed dedicated attention. Touch movement, auto-fire, a dash button, and a separate mobile HUD keep the controls and incoming attacks readable.
Accomplishments I’m proud of
- Building a complete playable loop from mutation selection to combat, results, and sharing.
- Giving the boss a recognizable personality and visual identity.
- Turning close dodges into an offensive advantage through Overdrive.
- Keeping creator credit attached to shared bosses.
- Passing 23 automated tests, alongside browser checks covering combat, persistence, challenge sharing, and touch controls.
What I learned
Increasing difficulty is more interesting when players gain something to master. Clear warnings, deliberate movement, and useful feedback create a stronger fight than simply adding more projectiles.
I also learned that community participation works best when its consequences are visible. A viewer’s idea, a credited attack, or a well-timed shield can become the moment everyone remembers.
What’s next
The next milestone is validating and recording the complete flow with real Twitch accounts: viewer proposals, authenticated Codex generation, live power voting, and a two-channel raid handoff.
These integrations are implemented, but have not yet been verified end-to-end against live accounts. Current integration tests use a controlled generator and injected chat events.
After that, I want to expand the boss roster, add more attack primitives, and use human playtesting to refine difficulty