[ SOLO · DESIGN + FULL-STACK · 2026 ]

An AI that runs your whole tabletop night — eight games, one engine.

One engine, eight live games: AI-narrated storytelling, Cthulhu investigation, a flagship murder mystery, D&D, lateral-thinking soup, truth-or-dare, an original card game, and clocktower-style social deduction. The AI generates and locks the content, plays autonomous agents, fills empty seats, narrates in neural voice, and runs it all in real-time multiplayer. Shipped as a real product: bilingual, voiced, monetized.

Play the live game → View source ↗
→ ROLESolo — product, design, frontend, backend, prompt/LLM orchestration, payments, deploy.
→ TYPESelf-initiated, shipped live product at mystnight.com — with real users, accounts and payments.
→ STACKNext.js 14 (App Router) · Supabase (Postgres + RLS + Realtime + Auth) · DeepSeek / OpenAI (text) · OpenAI & Azure Neural TTS (voice) · Stripe · Vercel.
→ MODESStoryteller · Cthulhu investigation · murder mystery · D&D · lateral-thinking soup · truth-or-dare · an original card game · clocktower-style social deduction — eight modes, one engine.
2
players to start
8
game modes
live-generated cases
2
languages (auto)
1
solo dev
▶ PROMO · 60 SECONDS

See a night come together.

A short look at the product: a case generated from nothing, two players at the table, and the moment the accusation lands.

DESIGN · SYSTEM

One token-driven design system — eight game modes, Figma to code.

Every screen runs on the same system: a three-tier variable set (primitives → semantic → scale), a component & pattern library, and a Tailwind preset that mirrors the Figma tokens 1:1 — change a colour in Figma and it's the same class in code. One cosmic-horror, serif, dark aesthetic across storytelling, Cthulhu investigation, D&D, a murder mystery and an original card game.

MystNight design tokens — brand colours, semantic tokens, and the serif type scale
Foundations — colour variables (primitives → semantic) + the serif type scale.
ink fog parchment blood eldritch amber green sky
16primitives
35components
9patterns
8screens

Variables → semantic tokens → components → patterns → full screens, all wired to a Tailwind preset + a starter React library & Figma Code Connect, so design and code never drift.

Open the system in Figma →

The screens below are rebuilt 1:1 from the live app's React components — the same layout, classes and token colours that ship on mystnight.com, rendered here as live DOM (not screenshots). Hover to enlarge, click to open full-size.

Home · mode selector
Cthulhu · investigation dashboard
Storyteller · narrated reading
Midnight Cat Curse · card table
Original card art — live components, drawn as line vectorshover to inspect · drag to deal ✋
THE MODES · ONE ENGINE, EIGHT GAMES

Eight different tabletop nights — one AI engine underneath.

Every mode below is live at mystnight.com and rebuilt here 1:1 from the app's own components. They share one core — locked AI generation, real-time multiplayer, neural voice and a single design system — but each is a genuinely different game, with its own rules, its own screen, and its own reason to exist. Click any screen to open it full-size.

01 · STORYTELLER

An AI that writes and performs an original story1–6 listeners · narrated · synced

The AI plans a story, drafts several versions, scores its own work and keeps the best — then reads it aloud in a chosen neural voice with the current line highlighted karaoke-style and adaptive music under each paragraph, synced for two listeners.

Why it's strongIt doesn't just dump text — a plan → draft → self-score → keep-best loop rewrites until the story actually lands, then voice and music turn reading into a shared, cinematic moment.
Ahead of the curveMost "AI story" toys stop at a wall of text. This closes the loop with self-evaluation and synchronized multi-listener audio — generative content treated as a directed performance.
Skill it provesLLM pipeline design (planning, self-critique, best-of-N) · TTS/audio engineering · real-time sync.
02 · CTHULHU INVESTIGATION

A cosmic-horror RPG with a real ruleset1–6 investigators · SAN · dice · clues

An AI Keeper runs skill checks, sanity, multi-clue deduction and a ticking world clock; investigators act in turns and the world reacts to what they actually do — with scene art generated on the fly.

Why it's strongA real system, not freeform chat: dice, suspicion and clue-gating are adjudicated server-side, so the AI narrates but can't cheat or spoon-feed the answer.
Ahead of the curveIt fuses a deterministic tabletop ruleset with an LLM narrator — exactly the hard, current frontier of "LLM as a fair game master."
Skill it provesHybrid deterministic-engine + LLM design · structured output · generative image pipeline.
03 · MURDER MYSTERY · FLAGSHIP

A locked case, run by an AI host who can't spoil it1–8 players · 7 acts · autonomous suspects

Pick a script and a full, self-consistent case is generated and frozen — culprit, motive, timeline, evidence, red herrings — then played across seven acts with AI suspects who lie, scheme and try to frame you, ending in an accusation and a scorecard.

Why it's strongThe truth is locked at generation and the host model is hard-blocked from leaking it or speaking for real players; every suspect pursues its own agenda.
Ahead of the curveIt solves the "AI spoils its own mystery" problem with a server-side locked-truth architecture — trustworthy hidden information under an LLM. (Full breakdown below.)
Skill it provesTruth-integrity architecture · prompt safety & guardrails · live multiplayer state.
04 · D&D ADVENTURE

A dungeon crawl where the dice are real code1–6 party · d20 combat · levelling

A deterministic d20 engine owns attack rolls, AC, HP, spell slots, class features and levelling; an AI Dungeon Master owns the story, the NPCs and the consequences, with turn-based party combat.

Why it's strongDamage and dice are computed in code and never hallucinated, so combat is fair and legible while the model is free to do what it's good at — fiction.
Ahead of the curveA clean split of "math the engine owns" vs "fiction the LLM owns" — the architecture serious AI-GM systems are converging on.
Skill it provesDeterministic game-systems engineering · turn/initiative state machines · AI narration.
05 · LATERAL SOUP (海龟汤)

A secret the AI guards for fifty questions2–8 players · yes / no riddle

The AI holds a hidden solution and answers any player question only with Yes, No, Irrelevant or Partly, staying perfectly consistent as the table closes in — until someone solves it.

Why it's strongIt reasons about arbitrary, open-ended player questions for dozens of turns without ever leaking or contradicting itself — pure constrained-inference discipline.
Ahead of the curveA minimalist proof that an LLM can be a reliable, non-leaking oracle — the same primitive that powers the locked mysteries, distilled to its core.
Skill it provesConstrained reasoning & prompt design · lightweight real-time multiplayer.
06 · TRUTH OR DARE

A party deck that's never the same twice2–10 players · scalable intensity

Every prompt is generated on the spot, scaled to a chosen intensity and setting, and screened against an off-limits list the model must respect — a deck that adapts to the room instead of a fixed box of cards.

Why it's strongPrompts are fresh and context-aware — who's playing, where, what's forbidden — while the safety constraints are actually honored, not just suggested.
Ahead of the curveControllable, guardrailed generation for a social context: real personalization with hard boundaries, the thing casual AI usually gets wrong.
Skill it provesControllable generation · safety/guardrail prompting · fast, frictionless UX.
07 · MIDNIGHT CAT CURSE

An original card game, designed from scratch2–6 players · bluff & reactions

A last-cat-standing game of bluffing, reactions and targeted plays — original rules, original hand-drawn line-art cards, AI opponents that fill empty seats, and reactions synced live across the table.

Why it's strongIt's a from-scratch game design — rules, balance and card art — running on the very same multiplayer engine, proving the platform isn't just licensed formats.
Ahead of the curveOriginal IP plus AI seat-filling means a full table any time, solo or duo — social games that no longer need a crowd to start.
Skill it provesOriginal game & visual design · custom illustration · AI-agent opponents.
08 · BLOOD ON THE CLOCKTOWER

Hidden-role deduction, hosted by an AI that also lies1–8 players · night / day / execute

Good versus evil social deduction where an AI Storyteller runs secret roles, night actions, nominations and executions — and fills empty seats with believable hidden agents, so a famously host-heavy game runs on demand.

Why it's strongThe AI juggles secret information, phase order and real-time table talk while playing convincing deceptive agents — the hardest kind of game to host well.
Ahead of the curveIt makes a notoriously host-dependent genre playable any time, even solo, with the AI both refereeing fairly and deceiving believably.
Skill it provesComplex phase/state orchestration · hidden-information management · multi-agent role-play.
FLAGSHIP DEEP-DIVE · MURDER MYSTERY · 01 / PROBLEM

A murder-mystery night needs a group, prep, and a human host.

Scripted mystery kits are great — once you've bought one, scheduled six people, and read the host booklet for an hour. The format barely exists for two people on a weeknight, and once you've solved a script, it's spent.

→ 01 / PEOPLE

You need a full table.

Most mystery games fall apart below 4–6 players. Couples and remote pairs are left out.

→ 02 / A HOST

Someone has to run it.

A good session needs a game master who knows the truth, paces the reveals, and never breaks. That person can't really play.

→ 03 / REPLAY

One script, one solve.

Once you know the culprit, the box is dead. Fixed content doesn't replay.

02 / WHAT I BUILT

An AI game master that writes the case, plays the suspects, and runs the table.

Pick a script and the system generates a complete, self-consistent, locked case behind the scenes — truth, culprit, motive, timeline, evidence and red herrings — then runs it as a live multiplayer session.

→ HIDDEN CASE

Locked truth, every suspect framed.

For mystery scripts, every character — including the real players — gets a plausible motive and a secret. Only one is the real killer; the rest are deliberate misdirection. The truth is locked at generation and can't drift.

→ AUTONOMOUS NPCs

Suspects that lie and scheme.

AI characters act on their own goals: they accuse, defend, form alliances, and actively try to frame others — including you. They never speak for the real players, and they're influenced by what's actually said at the table.

→ REAL-TIME DUO

Two humans, one link.

Your friend clicks a link and joins instantly — no install, no signup. Private role info is isolated per player; public scene is shared and synced live.

→ VOICE + SCORING

Read aloud, then graded.

Optional neural TTS reads narration and each character in a distinct, gender-matched voice. At the end, every player gets a multi-metric scorecard on deduction, roleplay and key calls.

PRODUCT · THE LIVE TABLE

Mid-case: the suspects turn on you.

The shared scene streams to both players in real time; your private role and secret stay yours alone. AI suspects push their own agendas — here one of them is openly trying to frame you.

◗ MystNight ROOM · THE GLASSHOUSE MURDER LIVE ♪ VOICE ON ACT 3 / 7 · THE CONSERVATORY Rain ticks the glass roof. Margot lingers by the orchids, watching you a beat too long. A brass key lies where the body was found. CLUES FOUND · 3 / 4 THIS ACT ▸ BRASS KEY · ▸ TORN LETTER · ▸ ALIBI GAP YOUR ROLE · 真人 / REAL PLAYER Dr. Ellis Reed Botanist. Owed the victim money. SECRET · ONLY YOU SEE THIS You were in the greenhouse at 10 — but not for the reason they think. SUSPICION ON YOU · ▮▮▮▮▮▯▯▯ HIGH MARGOT VANE · SUSPECT (AI) "I saw Ellis by the greenhouse at ten. Ask him what he was really doing." — THE INSPECTOR TURNS TO YOU — YOU · DR. ELLIS REED "Examine the brass key for prints." INSPECTOR HOLT · HOST (AI) The key is warm — recently held. A monogram catches the light: not "E". Say or do something… ⚖ ACCUSE
03 / ENGINEERING

The hard part isn't a chatbot. It's making an LLM run a fair game.

→ STRUCTURED OUTPUT

JSON the game can trust.

Every turn returns strict JSON (narration, NPC lines, revealed evidence, private notes, act state) with schema enforcement, retry, and an OpenAI fallback when the primary model filters or truncates.

→ TRUTH INTEGRITY

The answer can't move.

The full case is generated once and stored server-side only; the host model is fed the secret file but is hard-blocked from leaking it, from spoon-feeding clues, and from impersonating real players (prompt + server-side filtering).

→ PACING ENGINE

Acts on a real clock.

5–8 AI-designed acts advance on a live timer (or host control), gated so you can only skip ahead once the act's key clue is actually found. Clues only surface on genuine investigation, never on autopilot.

→ SYNC + FALLBACK

Resilient multiplayer.

Supabase Realtime drives instant updates, with a polling fallback so invited (anonymous) players never desync — a bug I only caught by playing on two devices.

→ VOICE PIPELINE

Provider-agnostic TTS.

One endpoint, swappable between OpenAI and Azure neural voices by env flag, with per-character voice selection by gender and hash, a sequential playback queue, and storage caching so a line is never re-synthesized.

→ BILLING

Free to try, then credits.

Stripe credit packs for hosting, Google OAuth, a per-account credit ledger with webhook dedup, a free first game for every new account, and invited guests always free.

UNDER THE HOOD · CASE GENERATION

Before anyone joins, a whole truth is written — and locked.

Pick a script and the engine generates a complete, self-consistent case server-side: culprit, motive, timeline, evidence and deliberate red herrings. It's scored for solvability, then frozen — the host model is fed the secret file but hard-blocked from leaking it.

PIPELINE · GENERATE → REVIEW → LOCK (SERVER-SIDE, BEFORE PLAY) GENERATE FULL CASE · LLM STRICT JSON REVIEW SOLVABILITY SCORE RETRY IF WEAK 🔒 LOCK TRUTH FROZEN CAN'T DRIFT HOST MODEL GETS THE SECRET FILE — BUT IS HARD-BLOCKED FROM LEAKING IT — NO SPOON-FEEDING · NO IMPERSONATING PLAYERS case.locked.json · SERVER-ONLY · NEVER SENT TO CLIENTS { "culprit": "Margot Vane", "method": "digitalis in the nightcap", "motive": "the rewritten will", "time_of_death": "22:10", "suspects": [ { "name": "Dr. Ellis Reed", "guilty": false, "red_herring": "debt + greenhouse alibi gap" }, { "name": "Margot Vane", "guilty": "████", "secret": "████████████" ⟵ withheld from table { "name": "Inspector Holt", "guilty": false, "role": "host / pressure" } ], "acts": 7, "clue_gates": ["brass key","torn letter","monogram","confession"], }
04 / SHIPPED

Not a demo — a product with users, payments and a domain.

→ LIVE

mystnight.com

Custom domain, auto HTTPS, bilingual UI that auto-detects the browser language, link-preview cards, and a private live-analytics panel for the operator.

→ FROM 0 → 1

Whole loop, solo.

Design, full-stack build, LLM orchestration, voice, payments, auth, deploy, and growth — one person, end to end.

ENDGAME · REVEAL + SCORECARD

The accusation lands — then everyone gets graded.

When a player commits to an accusation, the locked truth is revealed and the case is walked back act by act. Each player receives a multi-metric scorecard on deduction, roleplay and the key calls they made.

CASE CLOSED · THE TRUTH The culprit was Margot Vane. Motive: the rewritten will. Method: digitalis in the nightcap, 22:10. The brass key was hers. ✓ SOLVED IN ACT 7 PLAYER 01 · DR. ELLIS REED (YOU) Score · 86 / 100 DEDUCTION89 ROLEPLAY82 KEY CALLS93 RED-HERRING RESIST75 VERDICT Read the monogram correctly and pivoted off your own alibi gap in time. Clean accuse. PLAYER 02 · MARGOT VANE Score · 71 / 100 DECEPTION92 ROLEPLAY86 COMPOSURE60 FRAME ATTEMPT48 VERDICT Nearly pinned it on Ellis — the warm key and the monogram gave the game away.