si-wiki-fetch — Deterministic Spirit Island Wiki parser

Why this exists

Previous attempts to fetch Wiki data via WebFetch (which runs an LLM summarizer over the page) produced hallucinated specifics — wrong costs, wrong speeds, wrong effect text, wrong elements. That data is unreliable.

This skill uses scripts/wiki-fetch.py to hit the MediaWiki API directly, get raw wikitext, and parse the {{PowerCardArticle|...}} / {{Spirit|...}} template fields deterministically. No LLM in the pipeline. No hallucination.

Hard Rules

1. Before citing any mechanical detail (card cost, speed, range, target, elements, effect text, innate thresholds, special rules): run this skill.
2. Never paraphrase card text in a way that invents mechanics. The parser output is the canonical form; copy it verbatim into chapters.
3. If the Wiki page is missing or the template fields are absent: say so explicitly. Flag [VERIFY physical copy] rather than inventing.
4. Cache respected: the script caches under .wiki-cache/ — re-runs are free. Only use --no-cache if a Wiki page has been updated recently.
5. Rate-limiting respected: default 1.5s between requests; exponential backoff on 429/5xx. Don't override --delay below 1s unless you have a reason.

How to invoke

# Single card
python3 scripts/wiki-fetch.py card "Crops Wither and Fade"

# Single spirit (spirit panel only — innates, special rules, unique-card names)
python3 scripts/wiki-fetch.py spirit "Shadows Flicker Like Flame"

# Spirit + all its Unique cards (most useful for authoring)
python3 scripts/wiki-fetch.py batch-spirit "Shadows Flicker Like Flame" --output-dir data/references/wiki/

# Every spirit in data/spirits.json (takes a while due to rate-limiting)
python3 scripts/wiki-fetch.py batch-all --output-dir data/references/wiki/

Output shape

spirit / batch-spirit output

{
  "type": "spirit",
  "name": "Shadows Flicker Like Flame",
  "expansion": "Base Game",
  "complexity": "low",
  "special_rules": "SHADOWS OF THE DAHAN ...",
  "innates": [
    {
      "name": "DARKNESS SWALLOWS THE UNWARY",
      "speed": "fast",
      "range": "1",
      "target": "any",
      "option": "sacred",
      "thresholds": [
        {"moon": "2", "fire": "1", "effect": "Gather 1 Explorer."},
        ...
      ]
    }
  ],
  "unique_cards": ["Concealing Shadows", "Crops Wither and Fade", ...],
  "unique_card_details": [ ... ]  // only if batch-spirit
}

card output

{
  "type": "power_card",
  "name": "Crops Wither and Fade",
  "spirit": "Shadows Flicker Like Flame",
  "card_type": "Unique",
  "cost": "1",
  "speed": "Slow",
  "range": "0",
  "target": "Any Land",
  "elements": ["moon", "fire", "plant"],
  "text": "2 Fear. Replace 1 Town with 1 Explorer. **OR** Replace 1 City with 1 Town.",
  "threshold": "No Threshold"
}

Canonical workflow when writing a spirit chapter

1. Run: python3 scripts/wiki-fetch.py batch-spirit "Name" --output-dir data/references/wiki/
2. Open the resulting JSON. Copy-paste into the chapter:

• Special Rules section: use spirit.special_rules.
• Innate Powers section: list each innate with name, speed, range, target, and each threshold's elements + effect.
• Unique Cards section: for each card in unique_card_details, write a block with cost, speed, range, target, elements, text.

1. Brett's At-a-Glance table: expansion = spirit.expansion; complexity = spirit.complexity; aspects still need physical-copy verification (the Wiki doesn't surface aspect lists cleanly).
2. Strategic-framing prose should be built from the mechanical facts, not in advance of them.

Known limitations

• Aspects list: not yet scraped. Wiki stores aspects on separate pages; the spirit-page template doesn't list them cleanly. Flag aspects with [VERIFY] until scripted.
• Cost/speed/target when card page lacks a field: the parser reports whatever the Wiki has; missing fields come through as null / unset.
• Element icons in text: the parser replaces {{fear}} → "Fear", {{explorer}} → "Explorer", etc. See TOKEN_MAP in the script for the full list.
• Adversaries, scenarios, event cards, fear cards, blight cards: not yet supported — the script currently has parsers for {{Spirit}} and {{PowerCardArticle}} templates only. Extensions welcome.

When to pull fresh data

• A spirit/card recently had errata → use --no-cache on that specific page.
• You're writing a chapter for the first time → batch-spirit will cache everything, safe to rerun.

Never use WebFetch for mechanical data

If you find yourself reaching for WebFetch("https://spiritislandwiki.com/..."): stop and use this skill instead. WebFetch runs LLM summarization; that's how the original hallucinations happened.