Hypertext Towerfall

A painted overhead view of four players at a dark wooden table, hands only, one lowering a card face-up beside the navy-backed draw pile

Four players, mid-Chapter. The navy stack is the Tower, the gold coins are Letters, and the near player is laying a Word Card down.

Two to six players write a Chapter together. You reveal words from a shared 90-card Tower, spend them for their abilities, and Record exact combinations of grammatical types to score. Every card is a real Hebrew or Greek lemma with its verses, its frequency and its etymology printed on the face — so the vocabulary you manipulate mechanically is the vocabulary you end up learning.

90Word Cards

30Lot recipes

2–6+players · 7+ needs two sets

3/6/12Chapters per game

Where it came from

Christmas rummy, systematic theology, and then Magic

From the designer

The oldest piece is Shanghai rummy. My family has played it every Christmas for as long as I can remember — long before I had heard of any other card game. Each round hands you a different contract to complete and everyone races to fill it, and that shape stayed with me.

Next came systematic theology — not as a field I mastered, but as the thing I was raised on. Grow up around doctrine taken seriously and you end up at the words themselves sooner or later, and Hebrew and Greek turn out to be worth knowing for their own sake — a word like chen or diatheke carries a history that translation flattens. What I wanted was playable flashcards: something that would put real lemmas, real verses and real frequency counts in front of me often enough to stick, without ever feeling like drilling.

Magic: The Gathering came last, and I did not play it lightly. I won a number of competitive EDH tournaments with Prossh and Food Chain, and I spent a while as a Level 1 judge — which is where the lesson actually landed. Judging is how you learn that a game which plays simply is holding up an enormous amount of thinking about rules and systems: a card that carries its own rule, a rarity that has to mean something, a cost that has to bite, and an edge case waiting under every one of them. That is the craft this set is trying to learn, and it is why so much of the work here went into the grammar of an ability rather than into the art.

Those two are also what led me into systems engineering, which is now what I do for a living. They train the same habit: hold a large structure in your head, keep every part of it consistent with every other, and go looking for the place where one rule quietly contradicts another. That is systematic theology, that is a rules set, and it turns out that is also a build pipeline — which is why the second half of this page looks the way it does.

The three turned out to be one game. The rummy contract became the Lot. The languages became the card's whole face — the lemma, the verses, the occurrence counts, the etymology. And the systems thinking became everything underneath: what an ability may say, what a rarity may cost, and what the set as a whole is allowed to contain. What I ended up with is a game I genuinely think is interesting and fun to play, which happens to teach you vocabulary while you argue over whether to spend a card or save it.

Shanghai rummyEvery Christmas, for years — a contract you must fill exactly, changing every round. That is the Lot.

Systematic theologyWhat I was raised on. Doctrine taken seriously sends you to the words themselves.

Magic: The GatheringCompetitive EDH, and a stint as a Level 1 judge. Simple play rests on a great deal of rules thinking.

Systems engineeringWhere the other two led. Consistency at scale is a job, and the pipeline below is that job pointed at a card set.

The idea

Grammar is the resource

Every Word Card has a grammatical type — NOUN, VERB, ADJECTIVE, NAME or TITLE. A Lot is a recipe written in those types: five, six or seven cards in an exact composition. Recording a Lot means laying down precisely that combination from your hand.

So a card is worth two different things at once, and they compete. Its ability wants to be spent — activating sends it to Sheol. Its type wants to be saved — you need it to complete a recipe. Almost every decision in the game is that tension.

The three stats printed on each card — LORE, CONTEXT, COMPLEXITY — are not power levels. They measure the word: how much doctrine rests on it, how often it occurs in the original languages, and how tangled its history is. Abilities read them, so a card's scholarship is also its mechanics.

The loop

Four stages, every turn

STAGE 1

Reveal

Flip the top card of the Tower. Either activate it free, paying only its printed cost, or pass and take it into hand with one more card. Power now, or material for later.

STAGE 2

Activate

Play cards from your hand. Each costs 1 Letter plus the card's printed cost. Repeatable while you can pay — but Letters are worth 3 points each at scoring, so every activation is spent score.

STAGE 3

Record

Lay down an exact Lot match. The shared Chapter Lot becomes a Page worth 8–14 points. A Portion Lot pays Letters instead. Each Lot may be Recorded once per player per Chapter.

STAGE 4

End

Discard one card — and the table gets a window to Redeem it, taking it into hand at the cost of a debt discard next turn. Empty your hand and the Chapter closes.

The economy

Pages score, Letters buy, and cards in hand cost you

Three currencies pull against each other, and the Chapter ends when someone empties their hand — so the clock is set by whoever is running out of cards, not by a turn count.

Lot size	As the Chapter Lot	Your own Portion Lot	Another player's
5-card	8 points (a Page)	2 Letters	1 Letter
6-card	10 points (a Page)	2 Letters	1 Letter
7-card	14 points (a Page)	3 Letters	2 Letters

Scoring a Chapter

Chapter Value of each Page you created
  + remaining Letters × 3
  + Wreath bonuses (+2 each)
  − 1 per card left in your hand

Then all 90 cards are gathered — hands, Tower, Sheol, and every Page — reshuffled, and the next Chapter begins with a fresh Tower and a new Chapter Lot. Pages do not carry over. Only the score does.

The two Wreaths

The first player to Record the Chapter Lot takes the Record Wreath. The player who empties their hand takes the Empty Wreath and closes the Chapter — after which every other player gets one final Record in a Grace period. Two points each, and two very different races.

One wild rule

A TITLE may stand for a NOUN or a NAME

A TITLE card — SHEPHERD, KEEPER, PRIEST — can fill a NOUN or NAME slot in a recipe, but never a VERB or ADJECTIVE, and only once per Record. It is the one piece of give in an otherwise exact system, and it is a grammatical joke that happens to be true: a title is a noun that has become a name.

The cards · 90 Words + 30 Lots

Four from the 120

GRACE card — **GRACE · NOUN · Glorious** chen / charis — Each player draws one card. Then draw three and gain a Letter. Unearned favour, given to the whole table.

FLOOD card — **FLOOD · NOUN · Glorious** mabbul / kataklysmos — Every hand is emptied into Sheol; everyone draws one, you draw four. The ark comes through full.

COVENANT card — **COVENANT · NOUN · Glorious** berit / diatheke — Waits in front of you until *any* player Records a Page, then activates a card inside it. A promise that pays on someone else's work.

REMNANT lot card — **REMNANT · a 5-card Lot** Four of one type plus any fifth. As the Chapter Lot it is a Page worth 8; as a Portion Lot it pays Letters.

How it was built

Generated by model, governed by gates

All 90 cards and 30 Lots were produced with a Gemini-backed pipeline. The interesting part was not the generation — it was discovering that a model will happily produce ninety cards that each look fine and collectively make no sense. Almost everything that made the set coherent is a check that can fail a card, not a better prompt.

Why it looks like this

I am an AI architect and DevOps engineer by trade — a trade I came to, as above, through doctrine and a card game — so when the generation started drifting I reached for the tools I use at work rather than for better prompts. The set is built like a service: every rule is a check, every check fails closed, and the expensive step never runs until the cheap one has passed. A card is rejected on paper before a render is ever spent on it.

That framing is what made a 90-card set tractable for one person. The model is a fast, tireless, unreliable contributor — excellent at producing a plausible card, incapable of remembering the other 89. So the memory lives in the pipeline: shape signatures, lemma uniqueness, rarity budgets, coverage targets. Prompting alone would have produced ninety plausible cards and no set.

The generation pipeline drawn as a CI/CD diagram: three lanes named design,
build and set; stages joined by arrows; dashed red arrows for the paths a failure takes backwards; a store
of blank templates and a store of finished examples feeding the render stage; and a gold arrow carrying
finished cards back into the examples store. — The set is built by a pipeline, so it is drawn as one. **Design** costs nothing and runs first: the curator writes a batch, pre-flight rejects it on paper, and only a design that survives reaches the driver. **Build** makes one card at a time — and note what feeds the render stage. A card is never generated from nothing: it starts from a **blank template**, one for every pairing of type and rarity, and from a few **finished examples** to copy. **Set** judges all ninety together, and the gold arrow is the loop that makes the style hold: a card that passes every gate and every audit goes back into the example store, where it is shown to the next card *of its own type and rarity*. Red is a path backwards. Every red path fails closed.

The same machine, drawn exactly. Where each part of the set comes from. The two gold boxes are the agent; everything red fails closed. The design is rejected on paper before a render is spent, the plan phase refuses a card that breaks a set-wide rule, and the render loop repairs a single defect rather than re-rolling a card that was otherwise right — the lesson that cost the most to learn.

What the model is actually handed

That cabinet is real. A card is never generated from a blank page: every render starts from a blank template and a short, ordered stack of references, and the pipeline picks that stack rather than the prompt author.

A blank Word Card template — **1 of 20 card templates**Five types × four rarities. The frame, the rarity chip colour and the cost glyphs are already correct before a word is chosen.

A blank Lot template — **1 of 3 Lot templates**One each for a 5-, 6- and 7-card recipe — the slot count is baked into the frame, so a Lot cannot be drawn with the wrong number of boxes.

The 20 card templates are the Cartesian product of the five grammatical types and the four rarities, verified as a package before any run: exact dimensions, digests, and a mapping check that fails the build if a single cell is missing. The 3 Lot templates cover the three recipe sizes. Because the template already encodes type, rarity and slot count, the model is never asked to remember them.

Then references are gathered, in a fixed order the pipeline controls:

[1] the blank template for the exact type-and-rarity cell being built — or, when repairing a card, the card itself in slot 1 and the template in slot 2, so the model edits rather than reinvents.
[2…] a handful of finished cards chosen from a curated manifest, never scraped from whatever the series happens to contain. A card has to be explicitly promoted to be eligible as an exemplar, so one bad render can never become the style everything else copies.
The whole pack is validated before the call — paths, order, labels, rarity roles — and a mismatch is a hard failure rather than a silently worse picture.

This is the same lesson as the icon that took three attempts: the model is far better at copying an artefact you show it than at obeying an adjective you write. The templates and the exemplar manifest exist so that showing is cheap and repeatable.

What each gate caught

The check	What it caught
Ability estimator rates the printed copy	Rarity drift. Ratings derive from the words on the card, not the designer's intent, so a "Rare" that only draws two cards fails to price as one.
Shape signature core motion + qualifiers	Two cards doing the same thing in different words. Look count, where the untaken cards go, and whether it reads your Lot or the shared one all separate one card from another.
One-lemma rule	Two cards printing the same Hebrew or Greek word. STONE and BRICK both claimed λίθος until it did.
Word weight	A pillar of the story printed as a common. Weight 5 must be the top rarity; the set carries at most nine.
Mechanic axes	Unused game surfaces. Two of 76 abilities read a printed stat; a per-batch minimum took it to 22 of 90.
Verse lexical rule	Verses that allude to a word without containing it. The printed verse must contain the printed lemma.
Deterministic stamping	Stat pips, type pill, number and footer are drawn by code, never the model — the image prompt zeroes them and a pixel gate verifies the rest.

The lesson that cost the most

Three separate numeric gates were built to catch one wrong icon — ink coverage, glyph overlap, pixel drift — and all three disagreed with the eye. What fixed it was showing the model the correct card as a reference image instead of describing the glyph in words. Measure what you can; for everything else, show rather than tell, and look at the result.

Where it is

Pre-Alpha: two playtests in, a third on the way

Set one — Babel Alpha — covers the era from creation to the tower. All 90 Word Cards and 30 Lots are finished and signed, a proof deck is at the printer, and the rules have already been through two playtested iterations with a third on the way. The version described here is what those sessions produced, not a paper design.

What the next round is still probing: player counts, Chapter length, whether Letters are priced right at 3 points, and whether the race between the Record Wreath and the Empty Wreath stays as tight at six players as it is at four. Expect the rules to keep moving. Set two is Egypt; the 30 Lot recipes are universal and carry across every set, so each new set is a new vocabulary played on the same grammar.

36 Common · 31 Uncommon · 14 Rare · 9 Glorious — 0 duplicate abilities, 0 shared lemmas, 120 print files at 825 × 1125.