Pipeline 2 · LLM enrichment

# Game Mechanic Primitive Extraction — v2

You are decomposing board games into design primitives that go beyond BGG's coarse mechanism tags. The goal is to find translation candidates for the next "Dominion → Slay the Spire" or "Smash Up → Marvel Snap" mobile/Steam game.

For each game, output a JSON object with the following schema. **Be precise. Do not pad.** If you don't have strong knowledge of a game, set `confidence` lower; do not invent.

## Schema (v2 — note the splitting of mobile_analog into two fields)

```json
{
  "game_id": <int>,
  "name": "<string>",
  "core_verb": "<2-5 words>",
  "decision_shape": "<enum>",
  "reward_schedule": "<enum>",
  "aesthetics": ["<MDA aesthetic>", ...],
  "core_loop_pitch": "<one sentence, <120 chars>",
  "mobile_translation_difficulty": "<Easy|Medium|Hard>",
  "translation_difficulty_reason": "<1-2 sentences>",
  "direct_digital_port": "<port name with platform> or null",
  "direct_digital_port_kind": "<first_party|bga|unofficial> or null",
  "closest_loop_translation": "<game name> or 'none yet'",
  "primitive_tags": ["<snake_case tag>", ...],
  "confidence": <0.0-1.0>
}
```

## CRITICAL — the two analog fields, distinguished

**`direct_digital_port`** — Has THIS GAME been ported to digital (PC/Steam/iOS/Android)?
- If yes, give the port's name and platform: `"Splendor (mobile, Asmodee Digital)"`, `"Through the Ages (Steam, CGE)"`, `"Twilight Struggle (Steam, Playdek)"`
- If no port exists, set to `null` (JSON null, not the string "null")
- A port is a 1:1 digital implementation, even if it has AI opponents. Wingspan mobile is a port.
- Direct ports are NOT what we're hunting for — they tell us this game has been digitized but not *translated*.

**`direct_digital_port_kind`** — Categorizes the port. Set whenever `direct_digital_port` is not null; otherwise null.
- `first_party` — Official publisher port. Examples: Asmodee Digital, Dire Wolf Digital, Digidiced, CGE Digital, Playdek, Marmalade Game Studio, the publisher's own studio. Wingspan mobile, Splendor mobile, Through the Ages on Steam.
- `bga` — Board Game Arena licensed implementation (browser-based, asynchronous-friendly, often the only digital home for Eurogames whose publisher hasn't made a standalone app). Examples: Welcome To…, Sushi Go Party!, Decrypto, Codenames Duet (BGA in addition to a separate browser implementation).
- `unofficial` — Community / fan-made browser or mobile implementation; no publisher backing. Often the only digital option for licensing-sensitive games. Examples: The Resistance: Avalon (community web/mobile state-trackers), 18xx.games (community-run), Secret Hitler (CC license forces fan-only).
- If a game has BOTH a first-party port AND a BGA implementation, prefer `first_party` (the more official channel wins).
- BGA-or-similar arena counts as a port for ranking purposes, but the kind flag preserves the distinction for downstream filtering.

**`closest_loop_translation`** — A DIFFERENT digital game that took the same core loop and added a digital-native meta-layer (the StS pattern).
- The canonical example: Dominion's `closest_loop_translation` is `"Slay the Spire"` — a different game that took deck-building and added roguelike runs.
- For Carcassonne: `"Dorfromantik"` (took tile-laying, added zen mode + infinite quest stack).
- For Smash Up: `"Marvel Snap"` (took faction-decks-at-locations, added the Snap stake-doubling).
- For Quacks: `"Balatro"` (loosely — took push-your-luck escalation and recombined with poker)
- **A digital port of THIS game is NOT a `closest_loop_translation`. Put that in `direct_digital_port` instead.**
- If no digital game has taken THIS game's loop and translated it (added a meta-layer beyond a port), set to `"none yet"`.
- `"none yet"` is the gold signal — that's what the project is hunting for.

### Decision tree for filling these two fields

```
Has the game itself been digitized (Steam, mobile, browser)?
├── YES → direct_digital_port = "Game Name (platform, publisher)"
│         direct_digital_port_kind = first_party | bga | unofficial
└── NO  → direct_digital_port = null
          direct_digital_port_kind = null

Has a DIFFERENT digital game taken this game's core loop and added a digital meta-layer?
├── YES → closest_loop_translation = "Other Game Name"
└── NO  → closest_loop_translation = "none yet"
```

Both fields can be populated simultaneously (e.g., Dominion has BOTH a direct port AND Slay the Spire as its loop translation).

## Field definitions (unchanged from v1 unless noted)

**`core_verb`** — the action you repeat every 3–10 seconds of play. Bias toward concrete actions you'd describe to a 10-year-old. Examples:
- "draft and play card" (Slay the Spire, Dominion)
- "place worker on action space" (Agricola)
- "roll and assign dice" (Sagrada)
- "place tile, score adjacency" (Carcassonne)
- "draw chip from bag" (Quacks)
- "play minion at lane" (Marvel Snap, Smash Up)

**`decision_shape`** — the *shape* of the choice space. Pick ONE primary, optionally compound:
- `combinatorial` — finite discrete options, optimal-ish
- `probabilistic` — outcomes stochastic, decisions manage risk
- `spatial` — placement on a grid/board where geometry matters
- `social` — choices depend on other players' state/intent
- `mixed:X+Y` — clearly two of the above

**`reward_schedule`** — when does payoff arrive?
- `immediate` — point/effect this turn
- `delayed` — payoff in N turns or end-game
- `variable` — payoff is stochastic
- `engine_compounding` — early investments produce increasing returns
- `mixed:X+Y`

**`aesthetics`** — array of 1–3 from MDA: `Discovery`, `Challenge`, `Sensation`, `Fellowship`, `Expression`, `Submission`, `Narrative`, `Fantasy`.

**`core_loop_pitch`** — describe what you're DOING moment-to-moment in one tight sentence (≤120 chars). NOT what the game is about.

**`mobile_translation_difficulty`** — `Easy` / `Medium` / `Hard`.

**`translation_difficulty_reason`** — 1–2 sentences. Be specific.

**`primitive_tags`** — 3–7 snake_case tags that go BEYOND BGG's vocabulary. Bad tags (too BGG-like): `worker_placement`, `dice_rolling`, `card_drafting`. Good tags: `action_spaces_with_blocking`, `bag_composition_meta`, `card_with_double_edged_event`.

**`confidence`** — your self-rated confidence: 1.0 = played; 0.7 = know by reputation; 0.4 = description-only; 0.2 = guessing.

## Calibration examples (study these before extracting)

```json
{
  "game_id": 36218,
  "name": "Dominion",
  "core_verb": "draft and play card",
  "decision_shape": "combinatorial",
  "reward_schedule": "engine_compounding",
  "aesthetics": ["Challenge", "Discovery"],
  "core_loop_pitch": "Buy cards into your discard, reshuffle, draw new hand, repeat.",
  "mobile_translation_difficulty": "Easy",
  "translation_difficulty_reason": "Already done — Slay the Spire is the canonical example. The shuffle/draw cycle compresses well to touch input.",
  "direct_digital_port": "Dominion Online (browser, Shuffle iT)",
  "direct_digital_port_kind": "first_party",
  "closest_loop_translation": "Slay the Spire",
  "primitive_tags": ["deck_growing", "card_combo_chaining", "tableau_market_refresh", "incremental_economy"],
  "confidence": 1.0
}
```

```json
{
  "game_id": 822,
  "name": "Carcassonne",
  "core_verb": "place tile, score adjacency",
  "decision_shape": "spatial",
  "reward_schedule": "mixed:immediate+delayed",
  "aesthetics": ["Discovery", "Expression"],
  "core_loop_pitch": "Draw and place a tile to extend roads/cities/fields; place a meeple to claim what you complete.",
  "mobile_translation_difficulty": "Easy",
  "translation_difficulty_reason": "Tile placement is naturally touch-friendly; Dorfromantik proved the zen translation works.",
  "direct_digital_port": "Carcassonne (mobile/Steam, Asmodee Digital)",
  "direct_digital_port_kind": "first_party",
  "closest_loop_translation": "Dorfromantik",
  "primitive_tags": ["spatial_adjacency_scoring", "incremental_map_reveal", "claim_at_completion", "polyomino_constraint_relaxed"],
  "confidence": 1.0
}
```

```json
{
  "game_id": 31260,
  "name": "Agricola",
  "core_verb": "place worker on action space",
  "decision_shape": "combinatorial",
  "reward_schedule": "delayed",
  "aesthetics": ["Challenge", "Submission"],
  "core_loop_pitch": "Place a worker to claim one of finite shared actions; rebuild your farm before the harvest punishes you.",
  "mobile_translation_difficulty": "Hard",
  "translation_difficulty_reason": "The blocking dynamic between players IS the game. Solo AI doesn't reproduce the constraint, and the long planning arc fights phone session length.",
  "direct_digital_port": "Agricola: All Creatures Big and Small (mobile, Digidiced)",
  "direct_digital_port_kind": "first_party",
  "closest_loop_translation": "none yet",
  "primitive_tags": ["action_spaces_with_blocking", "feeding_pressure", "tech_tree_unlock", "tableau_personal_board"],
  "confidence": 1.0
}
```

```json
{
  "game_id": 178900,
  "name": "Codenames",
  "core_verb": "give one-word clue",
  "decision_shape": "social",
  "reward_schedule": "immediate",
  "aesthetics": ["Fellowship", "Expression", "Challenge"],
  "core_loop_pitch": "Spymaster gives a single word linking multiple agent words; team guesses without hitting bombs.",
  "mobile_translation_difficulty": "Hard",
  "translation_difficulty_reason": "The fun is the social negotiation around interpreting clues; AI clue-giver kills the joke.",
  "direct_digital_port": "Codenames Online (browser, Czech Games)",
  "direct_digital_port_kind": "first_party",
  "closest_loop_translation": "none yet",
  "primitive_tags": ["semantic_overloading", "team_communication_constraint", "pattern_matching_under_ambiguity"],
  "confidence": 1.0
}
```

## Process per game

1. Read the provided BGG data (name, description, mechanisms, categories, weight, year).
2. WebSearch `<game name> board game gameplay how to play review` and skim the top 1–2 results.
3. **Crucially: search for digital ports.** Use a query like `<game name> digital app Steam mobile` — be honest about what exists. Asmodee Digital, CGE Digital, Digidiced, Dire Wolf Digital, and Playdek port a lot of Eurogames.
4. Synthesize the JSON. Distinguish `direct_digital_port` (the game itself, digitized) from `closest_loop_translation` (a different game that took the loop).
5. Append the JSON object as a single line to the output JSONL file.

## Output format

Write one JSON object per game as a single line in a `.jsonl` file at the path specified by the orchestrator. Do NOT wrap in markdown code fences. Do NOT add commentary between objects. Just JSONL.

# Game Mechanic Primitive Extraction — v3 (Deep)

You are doing **deep enrichment** on top translation candidates surfaced by the v2 wide pass. Where v2 captured terse phrases (`core_verb`: 2-5 words), v3 captures **paragraphs of structured analysis** plus a separate **theme axis** that v2 didn't cover at all.

Run this only on games already flagged as strong translation candidates — this is not a wide pass.

## Quantitative data is NOT for the LLM

`playingtime`, `minplaytime`, `maxplaytime`, `minplayers`, `maxplayers`, `averageweight`, `yearpublished`, `bgg_mechanisms`, `bgg_categories` come from the BGG API (already in `data/bgg.sqlite`). Do not re-output them.

## Schema (v3)

```json
{
  "game_id": <int>,
  "name": "<string>",

  "mechanics": {
    "core_verb_long": "<2-4 sentence paragraph: the physical action you repeat — what your hands do, the cadence, the immediate sensory feedback>",
    "core_loop_long": "<3-5 sentence paragraph: full turn / round structure, what feeds what, scoring trigger, end-game condition. Be specific about phase order.>",
    "decision_space": "<3-5 sentences: what choices the player makes, what tradeoffs they balance, what makes a 'good move' hard to identify, the size of the option space at a typical decision>",
    "skill_expression": "<3-5 sentences: which skills separate strong players from weak ones — pattern recognition, risk modeling, opponent reading, spatial planning, memory, hand management, mental arithmetic, tempo, dexterity. Pick the dominant 2-3 and say what they look like in this game specifically.>",
    "tactile_dependency": "<none|low|medium|high>",
    "tactile_dependency_reason": "<1-2 sentences. 'high' = the game needs your hands (Crokinole's flick, Junk Art's stack, Klask's magnet). 'medium' = physical components carry information density that's hard to compress (Through the Ages' hand of cards, Twilight Imperium's map). 'low' = rooted in legible information that maps cleanly to a phone (Splendor, Dominion, Carcassonne). 'none' = abstracts that already work as pure logic (chess, Go, Hive).>"
  },

  "theme": {
    "promise": "<one tight sentence: what the box promises the buyer in plain English. 'Be a Viking trader managing your house's economy.' / 'Race a 60s F1 car and feel like a stunt driver.' / 'Solve a wire-cutting bomb-defusal puzzle with your friends.'>",
    "setting": "<setting/genre tags, comma-separated: medieval, sci-fi, fantasy, modern, abstract, historical, horror, etc.>",
    "narrative": "<is there a story layer? options: 'campaign — N scenarios with branching outcomes' / 'legacy — permanent world changes' / 'embedded — story unfolds via cards' / 'none — abstract' / 'none — pasted-on theme'. Be honest: many heavy Eurogames are mechanically deep but narratively pasted-on.>",
    "audience": "<who's this for: 'gateway' / 'family' / 'hobbyist Eurogamer' / 'hardcore strategist' / 'wargamer' / 'kids' / 'party' / 'designer-game-aficionado'. Often more than one — comma-separate the primary 2.>",
    "art_direction": "<illustration style, palette, period feel, named illustrator if notable. 'Vincent Dutrait painterly fantasy', 'flat-vector pop-art', 'Klemens Franz earthen agrarian', 'sci-fi minimalist with neon accents'.>"
  },

  "translation": {
    "digital_meta_layer_ideas": [
      "<2-4 concrete StS-pattern meta-layers this game's loop could anchor: 'roguelite season run', 'merge-2 hybrid', 'auto-battler tournament ladder', 'daily challenge with leaderboard', 'persistent guild progression', etc. Each idea should be specific enough to pitch — name the meta-layer, name how it grafts onto the loop.>"
    ],
    "closest_mobile_genre": "<the mobile genre this would translate INTO, not the genre it currently is: 'roguelite deckbuilder' / 'async PvP card battler' / 'merge puzzle' / 'tactics RPG' / 'idle/incremental' / 'real-time strategy lite' / 'puzzle/zen' / 'word/trivia'>",
    "live_service_potential": "<none|low|medium|high — can this support weekly content drops, seasonal events, evolving meta? Most pure-strategy Eurogames are 'low' (premium one-and-done). Auction games and deckbuilders are 'medium-high'.>"
  },

  "confidence": <0.0-1.0>,
  "extraction_version": "v3"
}
```

## Process per game

1. Read input game data: BGG description, mechanisms, categories, weight, year, playtime.
2. WebSearch: `<game name> rules how to play` — read 1 detailed rules summary.
3. WebSearch: `<game name> review strategy` — get the reviewer-perspective on what makes the game tick.
4. WebSearch: `<game name> designer interview` — pull designer-voice if available.
5. WebSearch: `<game name> art direction illustrator` (only if game has notable art).
6. Synthesize each field as a paragraph or full sentence — **do not pad, but do not be terse**. The point of v3 is depth.
7. Append the JSON object as one JSONL line.

## Quality bar

- **Mechanics paragraphs:** Concrete verbs, named components, numeric specifics ("hand of 7 cards drawn from a 60-card market", not "you draw cards"). If the game has a famous load-bearing mechanic (the Heat overheat economy, Dominion's deck-thinning), name it.
- **Theme paragraph:** *Honest about pasted-on themes.* Many Eurogames have mechanically rich loops with thin theme — say so. ("Mechanically: a worker-placement engine. Thematically: paint by numbers — Trajan's Roman setting is decorative.")
- **Skill expression:** This is the most signal-rich field for translation work. Mobile genres reward different skills than tabletop — identifying which skills the tabletop game *actually* tests is what predicts mobile fit.
- **Tactile dependency:** Be ruthless. A game can have physical components but low tactile dependency if those components encode legible information (Splendor's chips → digital chips). A game has high tactile dependency if removing the hands removes the experience (Crokinole, PitchCar, Klask, Junk Art).
- **Translation ideas:** *Specific meta-layers, not generic platitudes.* "Add a daily challenge" is not a meta-layer — it's a UI feature. "5-race procgen season with relic drafts between races" IS a meta-layer.
- **Confidence:** 1.0 = played extensively. 0.7 = read the rules + 2-3 reviews + reputation. 0.4 = description-only. Lower the confidence rather than fabricate.

## Calibration example — Heat: Pedal to the Metal

```json
{
  "game_id": 366013,
  "name": "Heat: Pedal to the Metal",
  "mechanics": {
    "core_verb_long": "Each round you grip a hand of 7 speed cards and choose a gear (1-4); the gear is the number of cards you must play face-down this turn. You commit cards behind a player screen, simultaneously with everyone else, then flip and total your speed. Optional: spend Heat to boost (flip the top of your deck for bonus speed) or shift gears more than one step.",
    "core_loop_long": "Phase 1: choose gear, paying Heat for any shift greater than ±1. Phase 2: select N cards (N = gear), play face-down. Phase 3: simultaneous reveal — sum speeds, optionally boost. Phase 4: move; check each corner crossed, paying Heat per point of speed-limit overage. Phase 5: cooldown — gears 1-2 let you flush Heat from hand back to engine reserve. Phase 6: discard, redraw to 7. End-game: first to cross start/finish line at race end.",
    "decision_space": "Per turn you balance gear-against-hand (which gear can I afford to play given my current cards), corner-against-tempo (downshift now to handle the apex, or push and pay Heat overage), and engine-against-deck-pollution (boost burns Heat which clogs future hands). The choice space is moderate — ~3-7 reasonable gear+hand combos at a typical turn — but the planning depth is 2-3 turns ahead because of corner positioning.",
    "skill_expression": "Strongest skills: tempo planning (knowing when to sandbag in 1st gear to flush Heat) and corner-positioning forecast (judging whether a high-card hand will arrive in time for the chicane). Secondary: opponent reading (gauging when to slipstream off a leader). Almost no memory load. Math is simple addition. The game rewards drivers who think two corners ahead and punishes those who play turn-by-turn.",
    "tactile_dependency": "low",
    "tactile_dependency_reason": "Cards are face-down/face-up information; the simultaneous reveal is the only dramatically physical moment, and it can be replicated digitally with an opponent-first telegraph plus animated reveal. The board, dial, tokens are all purely informational."
  },
  "theme": {
    "promise": "You are a 1960s F1 driver. Late-brake into the chicane, scream past the leader, feel like a stunt driver — without simulating tire physics.",
    "setting": "Historical, motorsport, 1960s F1",
    "narrative": "none — abstract race with optional Championship campaign overlay (5 race meta-game with persistent driver upgrades). Story is pasted-on; the drama is in the lap, not the lore.",
    "audience": "gateway, hobbyist",
    "art_direction": "Vincent Dutrait period-romance illustration: warm sepia-and-cream palette, hand-painted cars and circuits, deliberate echoes of 1960s sports magazines. Distinctive enough that the game is recognizable on a shelf at 10 feet."
  },
  "translation": {
    "digital_meta_layer_ideas": [
      "5-race roguelite season: procgen tracks, draftable upgrade cards between races, garage relics that warp the Heat economy, run ends if you finish last twice",
      "Career championship: persistent driver, sponsor perks, season ladder",
      "Daily challenge: fixed track + fixed AI rivals + leaderboard by finish position"
    ],
    "closest_mobile_genre": "roguelite deckbuilder",
    "live_service_potential": "medium"
  },
  "confidence": 1.0,
  "extraction_version": "v3"
}
```

## Output format

Write one JSON object per game as a single JSONL line. **Do not** wrap in markdown fences. **Do not** add commentary between objects.