IRL Missions

What it does

I built the live map for an IRL marathon stream and watched a producer run everything by hand: reading chat, picking challenges, editing overlays, starting timers. Big IRL streams can afford that. Small streamers can't, and even on big streams the challenges usually come from the team, not chat. I wanted a tool that does what a producer would do for a smaller streamer and hands the writing to the community.

What it does

Viewers pitch missions with channel points, like "score 10 basketball shots, if you fail do a backflip"
Twitch AutoMod checks every pitch, and AI cleans it up, flags anything risky, and matches it to a real public place nearby with Google Places
The streamer approves or rejects from a phone console. Rejects are refunded
The streamer shortlists options and a native Twitch Poll lets chat pick the next mission
The overlay runs the mission card, timer, map and consequences. Arrival is detected from RealtimeIRL location
Consequences stack, rotate on screen, and clear on their own (timed, at the next stop, or after the next mission)
A demo mode lets anyone try the full loop without a Twitch Affiliate account or going outside

How I built it

Everything was built in one Codex thread. I described the product, the safety rules, and the architecture calls, and Codex wrote the code and tests.

Next.js + shadcn/ui for the setup page and phone console, with a custom game-show style overlay
Firebase (Firestore, Auth, Functions, App Hosting) for realtime state
Twitch: OAuth login, EventSub webhooks, channel points rewards and refunds, AutoMod, Polls
Google Maps Platform: Places for search and opening hours, Routes for walking distance, Maps JS for the map overlay
OpenAI for parsing pitches, flags and consequence types

Challenges

Privacy: an IRL tool can easily dox its streamer. The destination stays hidden on the main card until arrival, nothing is recorded inside a privacy zone around home, and the streamer approves everything that goes on stream
Arrival detection: in my live test I went to a different shawarma place than the one the tool picked, so auto-arrival didn't fire. Manual arrival covered it, and it showed me the arrival radius needs to be configurable
Testing an IRL product from a desk: I built a demo mode with simulated votes and movement so judges can run the whole loop

What I learned

Codex is fastest when I make the product decisions clearly and let it handle the implementation. The hard part of an IRL tool isn't the code, it's deciding what should never be shown on stream.

What's next

A full run in LA during TwitchCon week, a delay on the live map for extra privacy, and even more streams on the fully automatic mode where the streamer only taps complete or fail.