Verb Forge — primitives, verb presets, and the screen that ties them together

Why, in one screen's worth of evidence

The Star Realms core-set probe (49 designs, imported clean, 15/15 games ending by Authority in ~21 turns) left 40 approximations. Every one traces to a verb the platform can't express: scrap this card, destroy target base, put the next ship you buy on top of your deck. Meanwhile "scrap" turned out not to need a mechanic at all — it is move of the SOURCE card into an out-of-game zone. The gap was never semantics; it was that verbs aren't composable.

Principle: the engine owns a closed set of primitives. Everything a designer sees — including attack — is data composed from them, with a per-game name. Nothing that a game means should live in a switch.

A · The contract

A.1 · Primitives (engine, closed)

Eight state mutators plus the value-source argument type (which already exists as ƒx). Adding one is a code change, on purpose.

A.2 · Verb preset (shipped catalog, assets/data/verb-presets.json)

A preset carries three things — the recipe, how the engine offers it, and a bot hint. The offer spec is the part with no precedent today and the one to design carefully: it is what the play table renders and what the bots enumerate.

{
  "id": "trash",
  "label": "Trash a card",              // default label; the GAME may rename it
  "category": "remove",                  // bot hint: attack | acquire | gain | remove | flow
  "slots": {
    "from":        { "type": "zone",  "required": true,  "label": "Trash cards from" },
    "destination": { "type": "zone",  "required": true,  "label": "Trashed cards go to",
                     "hint": "usually a shared, hidden, out-of-game pile" },
    "count":       { "type": "count", "default": 1 }
  },
  "offer": { "kind": "card_in_zone", "zone": "@{from}", "one_action_per": "card" },
  "steps_template": [
    { "do": "move", "cards": "$chosen", "to": "@{destination}" }
  ]
}

Offer kinds (closed): none (a single action, e.g. end phase) · card_in_zone · card_in_play · player · zone · affordable_card_in_zone (prices each candidate and drops what the actor cannot pay). The engine enumerates candidates, binds the choice as $chosen, and runs the steps.

A.3 · Per-game instance (game_meta['action_definitions'], unchanged shape)

"scrap": {
  "label": "Scrap",                       // the button in THIS game
  "enabled": true,
  "authored": { "preset": "trash",
                "params": { "from": "hand", "destination": "scrap_heap", "count": 1 } },
  "steps":  [ …compiled… ],               // recompiled on save; the engine reads only this
  "offer":  { …compiled… }
}

Same dual-key discipline as ƒx, conditions and today's custom actions: authored is the form's source of truth, the compiled halves are the engine's. Bindings (play_costs, uses_per_turn, phase_types, requires) keep working exactly as they do now, for built-in and custom alike.

A.4 · Presets we ship on day one

The current built-ins, re-expressed, plus the ones the probes proved missing:

B · The Verbs screen (Anatomy · Actions, rebuilt)

One page where every verb of the game lives — built-in or invented, all in the same list. Replaces today's split between a hardcoded checklist and a separate "Custom actions" block.

Row anatomy: enable checkbox · editable label (inline rename) · which preset it came from · phases where it is allowed (chips) · a plain-English summary compiled from the filled slots · inline validation warnings · Configure · overflow menu (duplicate, view compiled steps, delete).

Add verb opens the preset picker, grouped by category (Move & draw · Economy · Combat · Card state · Flow · Advanced/composer), each with a one-line description and an example. Choosing one opens Configure.

Configure is the slot filler — the same widget family the Jaipur actions already use, extended:

Below the slots, the same panel shows Bindings (cost, uses per turn, allowed phases, requires) — today a separate button — and a read-only compiled preview: the steps and the offer spec, so nothing is opaque. The raw-steps composer stays as the Advanced escape hatch.

C · Enable / disable / availability

Three distinct switches that today are spread over two pages and one hardcoded list:

Disabling a verb never deletes its configuration — it just stops being offered, so a designer can A/B a rule without losing the setup (and the version diff will show what the change did to the game).

D · AI autofill

Same shape as the strategy generator that already ships: one background call through the CLI worker, the game definition as context, a closed JSON answer, designer review.

• Per verb ("✨ suggest setup"): fills this preset's slots from the game's zones and trackers — "trash: you have no out-of-game zone; create scrap_heap (shared, hidden) and send trashed cards there".
• Whole game ("AI setup"): proposes the verb list for the game as briefed, each with slots filled, as a diff to review — nothing is written until accepted.

Rails, unchanged from the strategy work: the answer is validated against the closed preset/slot vocabulary; unknown ids are warnings, never silent; the designer sees every proposed value before it lands.

E · Migrating the built-ins (without breaking four games)

1. Ship the primitives + preset catalog; the old hardcoded verbs keep running.
2. Re-express ONE verb (attack_opponent → spend_pool_at_player) and run Starforge under both: same seeds, same stored profile. The version diff is the equivalence proof — endings, pacing and per-card usage must not move.
3. Repeat per verb, one commit each, each with its diff.
4. When a game's action_definitions covers every verb it uses, the engine stops consulting the hardcoded switch for that game (feature-flagged per game, so the cutover is reversible).
5. Delete the switch only when all four reference games are migrated and green.

Old games with no action_definitions keep working forever: an absent instance means "use the shipped preset with its defaults", which reproduces today's behaviour exactly.

Phases & sizing

~7 days. The payoff to measure at the end: re-run the Star Realms probe and watch the 40 approximations fall to the handful that are genuinely exotic (copy-a-card, per-turn play counters).

Risks & open questions