Skip to content

DEMO-RUNBOOK.md

Latest commit

 

History

History
87 lines (61 loc) · 7.83 KB

DEMO-RUNBOOK.md

File metadata and controls

87 lines (61 loc) · 7.83 KB
Raw

Gemini Display Kit — Demo Runbook

One-liner: A CLI + SDK that lets a developer scaffold, dev-loop, and deploy Gemini-powered games on Meta Ray-Ban Display in three commands. Built live at the Google I/O Hackathon today. Two demo games + one bonus example ship in the repo — all prototyped today and refactored to depend on the framework.

3-minute stage script

Timings assume a single operator + a live Pixel + a live pair of Display glasses. Speed dial is set to "Stage" (slowest) on any example UIs.

0:00 — Hook. "Smart-glasses just opened up. Building games for the Meta Ray-Ban Display is a stack of plumbing — bridge, SSE, deep-link, tunnel, every Gemini surface wired by hand. We built the framework that erases that stack. Three commands. Watch."

0:15 — Live scaffold. Terminal up, fresh shell.

npx gdk create demo-game --template adventure
cd demo-game && cp .env.example .env  # pre-edited with the key
npx gdk dev

Within ~10 seconds the terminal prints: bridge URL, tunnel URL, deep-link, a giant scannable QR. "That QR is the install. The bridge holds the Gemini key. The tunnel is HTTPS — required by the glasses."

0:45 — Phone scan → glasses. Scan with Pixel, Meta AI app opens → "Add Web App" → Add → glasses HUD lights up: rune-portal scene + three glowing doors. "That image was painted by Nano Banana."

1:00 — Pinch. Pinch the Neural Band. The HUD shows "the agents are painting your world…" → ~16s later a brand-new scene + three new doors. "Every door is a fresh image, generated on demand. Pure creative — not an image analyzer."

1:20 — Tour the framework while the next scene paints. Pull up the repo in the editor:

  • src/bridge/server.mjs — 350 lines. HTTP + SSE + static + per-game routes. Zero game logic.
  • src/tools/managed-agent.mjs — the wrapper that makes a managed agent a one-liner with persistent threads and trace capture.
  • templates/adventure/omni.config.mjs — 90 lines. This is the entire game. A defineGame call with two route handlers. The framework does everything else.

"The thing you just installed on the glasses was generated from this 90-line file."

1:50 — Cut to PulseBlade (second demo). Open examples/pulseblade/ in the editor side-by-side with the running game. "Same framework. Different mechanic. Flash designs the level structure in JSON, Lyria scores the backing track, a managed-agent director publishes a balancing note that streams to a judge console." Show the level designer prompt + the captured Lyria audio bytes + the director SSE event in the network tab. "Same bridge, same SSE, same tunnel. Different game."

2:30 — Pivot to PROMPT ARENA (the managed-agents money shot). Open examples/prompt-arena/data/traces/m_emberton.json in the editor. Scroll. "This is a real captured trace. The managed agent reads files, lists files, writes Python — there it is, code_execution_call, the actual code — and then runs it in its remote sandbox — code_execution_result, the stdout. That's the hatchery pattern. Eight monsters, eight agents, eight sandbox sessions. Each one tunes itself before fighting." Scroll through the bracket UI if time allows.

2:50 — The turn. "Framework, two demo games, one bonus example, all built today. Three managed-agent patterns — director, world-balancer, hatchery — proven in code. The framework is the moat: any developer can ship a glasses game tomorrow with the same one-liners."

2:55 — Provenance. "Public repo: github.com/sidkandan/GeminiDisplayKit. First commit at event start. Every line of the framework — and the example games — was written today. PROVENANCE.md has the file-by-file ledger."

1-minute video cut

Time Shot
0–8s hook + the gdk create / dev terminal moment, QR appears
8–20s phone scan → glasses HUD lights up → rune-portal scene
20–35s pinch → Nano Banana paints the next world (the visual wow)
35–48s the PROMPT ARENA hatchery trace — the managed agent writing & running code in a sandbox (the managed-agents prize shot)
48–55s PulseBlade quick clip — level JSON appearing, Lyria audio playing
55–60s tagline + "built today, public repo" callout

See docs/video-script.md for the shot-by-shot version.

Pre-warm checklist (do 10 min before slot)

  • .env file in examples/omni-odyssey/, examples/pulseblade/, examples/prompt-arena/ — each with a valid GEMINI_API_KEY
  • npm install already run in framework root AND in examples/omni-odyssey and examples/pulseblade
  • .venv already set up in examples/prompt-arena (Python venv + pip install -r requirements.txt if present, otherwise pip install google-genai flask python-dotenv)
  • One gdk dev rehearsal — ensure cloudflared brings up a tunnel within 10s (re-run if it fails the first time, the trycloudflare API is intermittent)
  • Pixel + glasses paired and the Meta AI app pre-opened
  • PROMPT ARENA: .venv/bin/python run_demo.py --hatch ran once so data/tournament.json is the frozen "golden" snapshot (cached strategies + sprites + audio)
  • OMNI-ODYSSEY: node scripts/prewarm-odyssey.mjs ran once so examples/omni-odyssey/assets/intro.mp4 (Veo cinematic) + opening-scene.jpg exist
  • Browser zoom on the laptop set to 125% so the terminal QR is scannable from 4 feet
  • Speed selector on PROMPT ARENA set to "Stage"

See scripts/demo-day.sh for the automatable parts of this checklist.

Q&A defense

Question Answer
"What did you actually build today?" The framework — src/, bin/, templates/. And the example games. Public repo, first commit at event start: git log --reverse --format='%h %ci %s' | head -3. The full file-by-file ledger is PROVENANCE.md.
"Is this an image analyzer?" No. See docs/anti-pattern-defense.md. The framework is provider-agnostic infrastructure; the games are creative (Nano Banana paints, Lyria composes, managed agents direct). Where we accept images (the optional Hint-giver pattern), it's as one input to a game-mechanic agent, not as "describe what's in this picture as a service."
"How is this managed agents, not just API calls?" Open examples/prompt-arena/data/traces/m_emberton.json — a real captured interaction.steps array including code_execution_call and code_execution_result. Server-side code execution + persisted state, not one completion.
"Who would use this?" Any developer with Display Web App access. The platform is brand new. There's no equivalent framework.
"Why one framework instead of multiple separate game submissions?" The framework is the multiplier-impact pitch. Each demo individually is one game; the framework is the means to ship hundreds.
"Why is Prompt Arena Python and the rest JavaScript?" Prompt Arena was prototyped earlier in the day as a standalone project; it's preserved verbatim as the bonus example for the code-execution trace evidence. A future refactor wraps its Flask server in an omni.config.mjs. The framework intentionally doesn't dictate language.

Reliability ladder (protect the 45% live-demo score)

  1. Primary: npx gdk dev in front of the judges — fresh scaffold, fresh tunnel, fresh QR. The CLI is the moat; show it working from zero.
  2. Tunnel fails (trycloudflare intermittent): re-run the same command — it succeeds on retry in our testing. Otherwise fall back to (3).
  3. Network dies during the live scaffold: switch to the pre-warmed examples/omni-odyssey/ running on gdk dev --no-tunnel against localhost on the operator's laptop — point at the running browser-tab HUD instead.
  4. Hard net failure: play the 1-minute recorded video and walk through examples/prompt-arena/data/traces/m_emberton.json in the editor (offline, captured artifacts).