How to playtest a card game solo — honestly

The slowest part of designing a card game isn't making cards. It's the two weeks between playtests, waiting for people. Here's what solo playtesting can genuinely tell you, where it lies to you, and how an AI opponent changes the loop.

The opponent bottleneck

Every designer knows the cycle: you tweak five cards on Tuesday, and you find out on the Saturday after next whether the tweaks worked — if the group shows up, if the game hits the table, if anyone remembers the old version well enough to compare. Call it two usable data points a month. At that rate a modest redesign takes a season, and most kitchen-table games quietly die in the wait.

The classic workaround is playing against yourself: run both hands, switch chairs mentally, try to forget what you know about the other side. It's genuinely useful for mechanical smoke-testing — does the turn structure work, does anything obviously break — and every designer should do it. But it has a known blind spot: you can't surprise yourself. You'll never make the greedy misread a stranger makes, never overlook the combo you designed, never value a card "wrong." Self-play tests your rules; it can't test an opponent's mind.

What an AI opponent fixes (and what it doesn't)

A language-model opponent sits in a useful middle spot: it isn't you, it reads your rules the way a stranger would, and it's available at 2am for the seventeenth game in a row. Concretely, solo AI playtesting is good at:

• Curve and tempo checks — is turn three always the same play? Does the aggressive line just win?
• Dominant-card detection — if the AI picks the same card every game, so will your players.
• Rules-text stress-testing — this is the underrated one. If the AI misreads a rule, a human will too. An AI opponent is a wording linter that plays games with you.
• Volume — ten games in an evening instead of two a month. Iteration speed is the whole point.

And honestly, what it's not:

• It won't feel the game for you — fun, tension, and table talk still need humans.
• Deep engine-builders and heavy hidden-information games will wobble; state-tracking over many intricate turns is where language models are weakest.
• It's a sparring partner for early iteration, not a replacement for the human playtests that come before publishing anything.

How the Cardwright Table does it

Cardwright Table is our version of the AI sparring partner, built to need nothing but what you already have:

1. Your rules, in plain English. No scripting language, no formal spec — write the rules the way you'd explain them to a friend. That text is exactly what the AI plays from (which is why it doubles as a wording test).
2. Your cards. Paste them one per line, or load a deck straight from the Cardwright Studio. Both players play mirror copies of your deck — the cleanest setup for balance testing.
3. Real shuffling. Your browser shuffles the decks (language models are famously bad at randomness); the AI referees and plays from the true shuffled order.
4. Flag anything illegal. When the AI misplays, one click replays the turn under a corrected reading — and you've just found a sentence in your rules worth tightening.
5. Share the table. Any game becomes a link; a friend who opens it plays your game against the AI immediately — no download, no account. Your playtesters no longer need to own anything or learn anything first.

It's free while in beta, and the whole game travels in the link itself — nothing is stored on our side.