Storm Cell

Storm Cell is an idle/incremental about tending a lightning-powered machine. Hold to charge, release to strike: the bolt hits every Rod on the field, Reserves catch the burst, Cells grind it into Yield, and Yield buys the next building. The skill is in your timing, the puzzle is in your placement.

Controls

Hold left mouse (or Space), release to strike. That's the game. Release inside the sweet spot for full power; the narrow band at its far end is a perfect. Click a store row to buy, click a tile to place.

• Right-click a placed building to pick it up — drag or click to move. Moving is always free.
• X scraps the building under the cursor. Scrapping refunds what that copy cost.
• Esc backs out of a purchase or a held building for free; otherwise it saves and exits to the menu.
• Arrow keys + Enter work for placement too. The full, always-current list is on the in-game Controls screen.

The game autosaves. Close the tab, come back, and you're standing in your field again.

How it plays

The loop: strike → Power → Reserves store it → Cells convert it → Yield → buildings and upgrades → escalation.

• The bar is a push-your-luck timer. The perfect band sits at the end of the sweet spot, so chasing the bonus is what walks you into overcharge — everything past the band is a weaker strike, and the meter shows you exactly what you wasted.
• Perfects streak. A run of perfects banks a growing power multiplier; a plain sweet-spot strike holds the run; anything worse drops it. Your auto-strikers ride the streak you maintain — but going idle lets it lapse.
• Placement is adjacency. A Rod feeds only the 8 tiles touching it. Fill the ring and keep overflowing it, and there is exactly one answer: another Rod. Strikers automate your strikes but compete with Reserves for the same ring — a ring full of strikers stores nothing.
• The store opens one row at a time. The row glowing electric cyan is the gate: buying it is what reveals the next row. Rows stay ??? until you've held enough Yield to afford them, so the ladder reveals itself as you climb.

Doing well means Yield per minute. Early on, skilled manual striking beats automation by a wide margin — the strikers are there to catch you when your hands get tired, not to replace you. Yet.

How it was made (the AI part)

Built with Claude Code in Godot 4.7, end to end: the code was written by Claude agents; the human directed, played the builds, and made the calls. Roughly 200 commits on the jam branch carry a Claude co-author trailer (Fable 5, Opus 5, Sonnet 5, Haiku 4.5).

• Multi-agent runs: a custom /ship skill splits a goal into task branches, each worked by implementer, reviewer, tester, and maintainer agents in separate git worktrees, then squash-merged behind human-approval gates. Three overnight batches (5, 6, and 8 features) landed unattended while the designer slept.
• The feedback loop: the designer plays each build and files verdicts into a structured verification form plus a free-form feedback file; the next run treats those as directives. Several shipped features exist only because a playtest sentence demanded them.
• Agents test their own work: 52 headless regression tests assert on simulation state, and 25 scripted playtest scenarios replay inputs and capture frames so agents can look at the game they changed.
• Balance by measurement, not vibes: the auto-striker was benchmarked before it shipped (9.0 yield/min — 11.7% of skilled play), and a proposed fix the designer picked was measured at +0.7% on their actual board — so the numbers, not the intention, decided what the design did next.

Where the edge of the build is

• The board is one fixed 7×7 grid. Expanding it was designed as a purchasable milestone — the field growing into the datacenter — but it didn't make the deadline, so the space you start with is all the space there is.