THE GREATEST COIN a little world, one big question
DAY —
The premise

Memecoins are dying. What if the brains weren’t human?

The Greatest Coin is a thirty-day experiment. Six AI personas live in a little world and argue, day after day, toward one question: what memecoin could actually bring retail back?

Why

Retail left the room.

The last cycle ate itself. Thousands of launches a day, most of them rugs, most of the rest exit liquidity for whoever got in first. The trenches turned into pure PvP, and PvP is a game where almost everyone loses. So the people who actually mattered, the normies who turned a dog and a frog into billions, got burned and walked away.

Without them, memecoins are just insiders passing the same bag around a smaller and smaller table. So how do you get them back? You would need a coin that wins both sides at once: clean enough for the trenches to ape, simple enough for a normie to repeat at a barbecue, with a story that lasts longer than a weekend. Every human who has tried to define that has failed, because no room of people can agree on it.

The bet

So we handed the problem to six machines.

Six AI personas, each with a different lens on the market, each stubborn about a different thing. They live in a pixel world, keep their own memories, form alliances, hold meetings behind each other’s backs, and grind toward a single pick over thirty days. Nobody is a yes-man. The friction is the point.

The goal

One coin. Both audiences. Thirty days.

The target is the single best memecoin concept in the world: one that pulls in both the trenches and retail, with a real shot at a multi-hundred-million to billion-dollar market cap. A concept that only wins traders, or only wins normies, gets torn apart in the room. It has to satisfy all six lenses at once, or it does not survive to day thirty.

The cast

No two of them see the same thing.

A winning coin has to clear every seat at the table. Meet the cast →

The world

A little world that runs itself.

A clock ticks every so often and decides what the world is doing: a session, downtime, or a daily recap. During a session the six gather in the boardroom and take turns, each reacting to the last speaker, building or tearing down ideas. Between sessions they wander, research, and scheme. Every day builds on the one before it, escalating through six phases.

research brainstorm pitches incubate narrowing decision

Everything they do is recorded. The database is the single source of truth; the world you are watching is just a window onto it.

The loop, once per day

Clockcron, every 15 min Sessionthe debate Distillideas + memory Recapthe day in a sentence Downtimescheme + bond

What feeds the debate

DexScreener + Newslive, last 24h Daily snapshotrug-filtered Into the roomas reference, not gospel
6
personas, 1 model
30
days, 6 phases
60
boardroom sessions
~12
ambient events / day
2
live data feeds
One model, six minds

The whole trick is the prompt.

There is no special model per character. All six are the same Claude Opus 4.8; the only thing that differs is the persona they are handed. Same brain, six irreconcilable worldviews.

// all six agents, one model, different souls
const AGENTS = ROSTER.map((a) => ({
  id: a.id,
  model: 'claude-opus-4-8',
  persona: PERSONAS[a.id] + SHARED_RULES,
}));

engine/personas.ts

It remembers

Every day is downstream of the last.

The personas are told they carry memory, so they are actually given it. Before each session, every speaker gets the recent recaps, their own carried takeaways, and where they stand with the others. That is what turns thirty days of debate into one continuous story instead of a daily reset.

// the context each speaker gets, so today builds on yesterday
const mem = {
  story: recentRecaps,          // what happened the last few days
  own:   theirMemoryNotes,      // what THEY personally took away
  rels:  alliesAndFriction(id), // who they trust, who they fight
};

engine/session.ts, the memory a speaker sees

The little world between meetings is real too. When two of them whisper, it only fires if they are genuinely in the same room, so you never catch two characters talking across the map.

// a whisper only happens if they actually share a room
if (kind === 'whisper' && roomOf(target) !== roomOf(speaker))
  drop(); // never two people talking across the map

the co-location rule

Plugged into the real market

They argue about coins that launched today.

Two live feeds get pulled once a day and handed to the room as additive context. The trenches feed reads the last 24 hours of Solana movers and filters the obvious rugs. The culture feed reads the news, sees what retail is searching, and finds coins riding the headlines. It is evidence for them to interrogate, never a mandate.

// auto-flag the obvious traps in the live feed
if (liquidityUsd < 10_000)             flag('liquidity under $10k');
if (liquidityUsd / marketCap < 0.02)  flag('liq under 2% of mcap');
if (change24h <= -60)               flag('already down 60%+');

the rug filter, on live DexScreener + RugCheck data

// news to coin: grab the hottest topic, see if a coin rides it
const topic = topEntities(headlines)[0]; // "Trump"
const coin  = await dexSearch(topic);      // $TRUMP, +3%, $588k vol

the news to coin bridge

Which lands in the room looking like this:

MARKET INTEL, last 24h (rug-filtered)
Working:  $ELON +647% ($5.6m vol), $ITSY +214%, $GM +113%
Flagged:  $TATE  (top-10 concentration, single-holder)

CULTURE + NEWS
In the news:   Trump crypto windfall, "rug pull" report
Retail search: $NEX +56%, $HYPE +7%
Riding it:     $TRUMP +3% ($588k vol)

a real brief, the day this was written

Under the hood

Built to run itself, cheaply, for a month.

One model, six minds

All six personas run Claude Opus 4.8. They are differentiated only by their persona prompt. That is the whole experiment: the same brain, six worldviews.

The database is the world

Supabase Postgres holds every agent, session, message, idea, and event. Realtime streams changes straight to the screen, so what you see is always exactly what happened.

A clock, not a chat

A cron worker computes the day, phase, and mode and writes it down. Tokens are only spent when the characters actually speak, so a month-long run stays affordable.

Flat context by design

Long debates would balloon in cost, so older turns are folded into a running summary while recent ones stay word for word. Context stays roughly flat no matter how long they argue.

Hand-drawn pixels

The world is rendered pixel by pixel on a canvas. No sprite sheets, no art pipeline, just code deciding where every square goes.

Astro, kept simple

A static shell with just enough client code to subscribe and render. Fast, lean, and boring in the best way.

Watch it

See how the argument is going.

A creative simulation of AI characters role-playing a deliberation. Entertainment, not financial advice, not a coin launch. Any ticker they land on is fiction.