The first balance pass measured each tower against the others. One Cannon did 1823 damage per 100 gold, an Archer 445 and a Mage 332, and a bot playing greedily built 97% cannons. The cause was a rule: a blast hit everything inside it for full damage. Now the target takes full damage and the rest take 65%, and the Mage breaks armour, so every tower has a job.

Region 1 is where new players quit, and the runs showed it: 25 of 40 simulated novices and 14 of 40 Archer-only starts died there. I had the curve rebuilt so each region ramps to big waves on 3 and 6, then wave 7 and the boss. Now nobody dies in region 1, a first-timer reaches region 2, and the skilled bot clears about a third of its runs.

The shop routine crashed on every tower offer, and supports were placed by comparing a range in squares to a distance in pixels, so the bot had never built a single support tower. The tools were fixed before the game was. Every run is deterministic, so a before and an after can be compared seed by seed.

The tests grew into a native Mac app. It lists every spell, tower, gem, crystal, animal and environment with its numbers, its 83 synergy links and a clip of each tower firing. A change is saved as a draft against the game's current value, and the game stays untouched until the drafts are applied.

After each boss comes the Shift: the left half of the world crumbles away with every tower on it, and a new region rises on the right. Where you build so your towers survive is half the game. Each biome does one thing to the whole board: Grassland strengthens towers the longer they stand, Desert sends sandstorms that cut range, Lava sets crystals alight and Snow slows them. Two halves of the same biome double the effect, and two different biomes combine into a new one.

The shop opens after waves 3 and 6 and after each boss. It sells spells that combine with each other, towers for your reserve rack and, from region 2, a voucher that lasts the run. You hold five spells and their order matters. The 28 spells fall into builds like archery, frost, pyromancy, precision, command, wealth and fortress. A new profile starts with Archers and six spells. Everything else unlocks through play, and an unlock joins the run in progress with a fanfare.

Shiftguard began as a 2D game in LÖVE. The move to Godot and 3D made the towers physical pieces: they stand on a shelf beside the board, and you pick one up, carry it and set it on a pad. I modeled the six towers and the gate by hand as flat-coloured toon shapes. Each grows through three tiers, then picks a specialisation.

Round one was oil and gouache: loose brushwork and canvas texture, with the background cut into flat geometric panels. Round two became the set: a clean illustrated subject on a two-tone patterned panel, one pattern per spell family so a build reads at a glance. Both were generated with a mix of image models, one card at a time from a shared prompt frame, with no text on the card face.

By hand: the brief (balanced the way Balatro feels, an easy start, then a climb), the tower and gate models, the card prompts and the pick of every card. With AI: the Godot code, the port from LÖVE, the tests, the balance bots and the editor, plus the card art and the hand cursors, which a script cuts to pixel art. Tools: Claude Fable, Claude Opus, Cursor, Gemini, Higgsfield, Figma Weave and Figma.

A playable beta with notes for testers: 7 towers, 28 spells, 4 biomes and a bestiary of 13 crystals, in Godot 4.7. Next is more of the same loop: more spells that change a build and more reasons to start another run.