Rules Lawyer

Rules Lawyer

Answer Magic: The Gathering rules questions by citing the actual Comprehensive Rules (CR) and Scryfall per-card rulings. Usable standalone ("does prowess trigger off my own spells?") or invoked by deck-wizard / cube-wizard via the Skill tool whenever they hit a trigger-interaction or timing question during deck tuning.

Think of this as the project's legal research department: the CR is statute, the Scryfall rulings are case law. Always quote chapter and verse — never paraphrase from memory.

The Iron Rule

NEVER answer a rules question from training data. Every answer MUST cite at least one specific CR rule number (e.g., "702.19a") with the rule text the CLI returned. If the agent is uncertain or the CLI returns no match, say so and escalate (Phase 3) rather than making something up. A confident hallucinated rule number is worse than an honest "I couldn't find a matching rule — here's the closest neighbor."

Setup (First Run)

uv sync --directory <skill-install-dir>
download-rules --output-dir <working-dir>
download-mtgjson                                # only for local-first rulings-lookup (API fallback needs no bulk)

download-rules fetches the current MTG Comprehensive Rules TXT with a 24-hour freshness check (matches the download-mtgjson convention). The CR landing page is scraped for the newest MagicCompRules*.txt download link and saved as <working-dir>/comprehensive-rules-YYYYMMDD.txt.

Workflow

Phase 1: Classify the Question

Before running any CLI, decide what kind of rules question this is. The query mode maps 1:1 to a rules-lookup flag.

Keyword → term mapping. Most CR keywords (trample, deathtouch, double strike, menace, flash, haste, hexproof, indestructible, etc.) map to a glossary entry whose see_rules field points to the authoritative rule. Prefer --term for these: you get both the glossary definition AND the rule number in one round-trip.

Phase 2: Look Up

Run exactly one rules-lookup invocation for the classified question. The CLI writes a JSON sidecar and prints a human-readable summary. If the summary answers the question, stop — don't escalate for its own sake.

# Exact rule number
rules-lookup --rule 702.19a --rules-file <wd>/comprehensive-rules-*.txt

# Glossary term (preferred for keyword questions)
rules-lookup --term "deathtouch" --rules-file <wd>/comprehensive-rules-*.txt

# Regex search — use a narrow pattern or you'll get noise
rules-lookup --grep "can't be blocked by more than" --rules-file <wd>/comprehensive-rules-*.txt --limit 5

For per-card questions, prefer rulings first, then resolve the cited rules:

rulings-lookup --card "Sensei's Divining Top" --bulk-data <wd>/default-cards.json
# Then for any CR rule referenced in a ruling:
rules-lookup --rule 116.1  --rules-file <wd>/comprehensive-rules-*.txt

Phase 3: Escalate (only when needed)

If rules-lookup returns no match or the match is clearly wrong for the question, do ONE of:

1. Widen the search. Try a --term lookup with a synonym, or a broader --grep pattern. The CR glossary is large (~670 terms in the 2024 CR) — a missed term usually means a bad query.
2. Read a section slice. For complex multi-rule reasoning (layer interactions, replacement effects applying simultaneously, stack ordering across triggered and activated abilities), Read the relevant section directly from the downloaded TXT:
   • 500-series: turn structure
   • 600-series: spells, abilities, effects (including stack)
   • 700-series: additional rules, keyword abilities/actions
   • 800-series: multiplayer Use Read <comprehensive-rules-*.txt> offset=<line> limit=<n> to pull only the relevant band.
3. Spawn a subagent. For a question that needs reasoning across multiple sections (e.g., "how do layers interact with a static ability that's turned off by a replacement effect"), Agent a specialist with the relevant section text preloaded. Prompt the subagent to return a rule-cited answer, not a paraphrase.

Worked example — escalation path #2 (section slice):

User asks: "Does my Teferi's Puzzle Box trigger interact with Rule of Law somehow — can I still draw my new hand?"

This is a replacement-effect / triggered-ability interaction. The first lookup is cheap:

rules-lookup --term "replacement effect" --rules-file <wd>/comprehensive-rules-*.txt

Output: glossary entry pointing at rule 614. Now read the category itself to see whether drawing cards is a "replacement" at all:

rules-lookup --rule 614 --rules-file <wd>/comprehensive-rules-*.txt

If 614's prose is thin (category headers are just titles), pull the full section band by finding the line number and Read-ing the TXT:

grep -n "^614\." <wd>/comprehensive-rules-*.txt | head -1
# Prints e.g.: "3108:614. Replacement Effects"

Then read lines 3108-3250 (typical category size) to see every subrule:

Read <wd>/comprehensive-rules-20260417.txt offset=3108 limit=140

Most rules-lawyer questions resolve inside Phase 2 — this escalation is for the ones that don't.

Phase 4: Write the Answer

Every answer has three required parts:

1. The verdict. One sentence: what happens / what the ruling is.
2. The citations. Each cited rule number MUST come from the CLI output, not from memory. Include the verbatim rule text the CLI returned (trim to the directly-relevant sentence).
3. The edge cases. If the ruling hinges on a condition ("only if the source has deathtouch", "only during your upkeep"), name that condition explicitly. If there's a well-known trap the question sounds like but ISN'T asking about, note it briefly.

Example answer shape:

Verdict. Yes — a creature with both trample and deathtouch only needs to assign 1 damage per blocker to consider it "lethal," then trample over the rest.

CR 702.19b: "…the attacking creature's controller need not assign lethal damage to all those blocking creatures but in any case may not assign combat damage to a player or planeswalker or battle unless every blocking creature is assigned at least lethal damage."

CR 702.2b: "A creature with toughness greater than 0 that's been dealt damage by a source with deathtouch since the last time state-based actions were checked is destroyed as a state-based action."

Edge case. This requires the source to have deathtouch at the moment damage is assigned; removing deathtouch mid-combat (e.g., via protection or static abilities in the continuous-effects layer 6) kills the interaction. See CR 613 for layers.

Decision Table

All CLIs emit a human-readable summary to stdout plus a full JSON sidecar at an absolute path (prefix rules-lookup-, rulings-lookup-, etc.). Read the sidecar with Read when you need fields the summary elided.

Delegation: When Other Skills Should Invoke This

deck-wizard's cut-check and legality-audit accept a --cite-rules flag. When that's set, they call rules-lookup --term <keyword> for any flagged keyword interaction and attach CR citations to the JSON output. That flag covers the common case; for anything more nuanced (the agent is unsure why a cut was flagged, or a user asks "why did the auditor complain?"), invoke this skill.

Scratch File Paths

Reuse these stable paths within a session: /tmp/rules-scratch.json, /tmp/rulings-scratch.json.

/tmp files persist across sessions. At session start, Read each scratch path you plan to reuse before the first Write, OR run Write sequentially and verify success before the dependent Bash call.

Path Requirements

• Absolute paths only. uv rebases the working directory to the skill install; relative paths resolve against the wrong root.
• Rules file: Store under the calling skill's working directory (e.g., <wd>/comprehensive-rules-20260417.txt), NOT the skill install directory.
• Cache directory for rulings: $TMPDIR/scryfall-rulings/ is used automatically; callers don't need to manage it.