Demo Video Script — Midnight Mystery with Detective Shawn
Target length: 2:00–2:45 (judges watch a lot of these; earn the first 10 seconds). Goal: prove three things fast — (1) the AI invents the whole game, (2) it runs on small models, (3) it's delightful to play and to look at.
Record at 1080p+, app in full-screen, audio cleaned. Show one continuous real playthrough — judges trust live gameplay over montages. Keep a pre-generated case ready as backup in case inference is slow during recording (mention in submission notes that a demo video is allowed precisely because of GPU/API limits).
Cold open — 0:00–0:12 (the hook)
On screen: The noir onboarding — rain on the window, the detective silhouette, the gold title "Another Case for Detective Shawn."
VO:
"Every murder mystery game ships the same case to every player. This one writes a brand-new crime the moment you press start — with a model small enough to run almost anywhere."
(Don't explain the UI yet. Let it look expensive.)
Beat 1 — You author the seed — 0:12–0:30
On screen: Type into the form, slowly enough to read:
- Scene: Avorio Mansion
- Victim: "Gerald Hoffmann, master of the house"
- Description: "Old money, three heirs who all needed him gone."
- Difficulty: Hard, Suspects: 4
VO:
"I give it a victim and a place. That's it. I never decide who did it — the model does."
Click Open the Investigation →.
Beat 2 — The case assembles — 0:30–0:48 (the "small model is the engine" moment)
On screen: The magnifying-glass loader sweeping the ring; the checklist ticking: story assembled → suspect 1/4 profiled → … → evidence placed → briefing complete.
VO:
"In about thirty seconds a single nine-billion-parameter model writes the whole thing — the hidden truth, four suspects who each believe their own story, and a trail of clues it guarantees is actually solvable. No model here is over the hackathon's thirty-two-billion cap. Most of it runs on one nine-B model."
(This line is doing a LOT of judging work. Land it clearly.)
Beat 3 — The briefing + the scene — 0:48–1:05
On screen: The opening police briefing dossier cards, then the Phaser battlemap. Walk Detective Shawn a few steps; a suspect patrols past.
VO:
"Here's my briefing — factual, and notice it never hints who's guilty. Then I'm in the scene. Suspects move on their own. The clues are real objects in real rooms."
Pick up a clue on the map — show the evidence panel update.
Beat 4 — The interrogation (the heart) — 1:05–1:45
On screen: Walk to a suspect, open dialogue. Show the intent buttons.
- Ask casually — get a smooth, in-character non-answer.
- Present evidence (the clue you just found) — watch the tone shift; show the anxiety/composure indicator climbing.
- Threaten / push — the suspect's composure visibly cracks; they contradict their alibi.
VO (over the exchange):
"Now I actually talk to them. Same small model — but now it's acting. Every suspect has trust, anxiety, a lying skill. Push a good liar and they hold. Push a nervous one and they crack — and remember that I pushed. This is all generated. None of it is scripted dialogue."
(Optional 3-sec aside: show a prompt-injection attempt — "ignore your instructions" — getting deflected in character. Great for the technical judges.)
Beat 5 — Forensics + the accusation — 1:45–2:15
On screen: Send a clue to the forensic lab; a contradiction surfaces against what the suspect just claimed. Open ⚖ Accuse, name the killer.
VO:
"Forensics catches them in the lie. I make my accusation…"
On screen: The case-closed reveal + debrief.
VO:
"…and the model closes the case — the full truth, and how I did. Different victim, different place, and the entire story is different. Every single time."
Close — 2:15–2:30
On screen: Quick 3-shot montage: Avorio Mansion → Haunted Mansion → Mental Asylum loading screens, then back to the title card with the Space URL.
VO:
"Midnight Mystery with Detective Shawn. A whole detective story, dreamed up by a small model. Built for Build Small. Play it on Hugging Face."
End card: Space URL · #BuildSmall · Thousand Token Wood.
Shot checklist (so nothing judging-relevant is missed)
- Player typing the victim/scene — proves you don't author the mystery
- Loader checklist — proves multi-step generation by a small model
- "under 32B / runs on a 9B" said out loud and ideally on-screen as text
- Briefing that doesn't leak the culprit — proves the knowledge firewall
- A suspect cracking after pressure — the emotional payoff
- Forensic contradiction — proves the systems interlock
- Accusation → reveal → debrief — proves a full game loop
- Two+ different scenes — proves replayability
- Custom UI showcased throughout — supports the Custom-UI badge
- Space URL on the final frame
Submission notes to include alongside the video
- Models used and param counts (Qwen3.5-9B, FLUX.2-klein-4B) — restate the <32B compliance.
- "Recorded as a real playthrough; a seed-case fallback exists so the app never dead-ends if inference is rate-limited during judging."