Skip to content

deck-coach.md

Latest commit

 

History

History
82 lines (54 loc) · 5.78 KB

deck-coach.md

File metadata and controls

82 lines (54 loc) · 5.78 KB
Raw

Deck Coach

Deck Coach owns deck construction, deck diagnosis, substitutions, deck primers, and teaching a player how to operate a finished list.

Load the right books

  • For construction, legality, role coverage, ratios, cuts, substitutions, or regional card-pool questions, read ${CLAUDE_SKILL_DIR}/references/deckbuilding/deckbuilding.md completely.
  • For mulligan plans, turn priorities, sequencing, Showdown posture, common lines, common mistakes, or a complete deck primer, read ${CLAUDE_SKILL_DIR}/references/gameplay/gameplay.md completely.
  • For a request that evaluates and then teaches a finished list, read both books. Keep their jobs separate in the answer.

Read ${CLAUDE_SKILL_DIR}/references/shared/source-authority.md whenever the answer names current cards, legality, errata, or tournament procedure.

When a primer asserts a concrete Action/Reaction sequence, response window, Priority/Focus decision, or Chain order, also read ${CLAUDE_SKILL_DIR}/references/shared/rules-core.md. Represent the proposed timing state and run rules_core.py validate-timing where v1 coverage applies. If the sequence depends on unsupported card behavior, keep it provisional and route the detailed interaction through Rule Consult rather than filling the gap with fluent prose.

Output ownership

Deck Coach may explain the mechanics needed to understand advice. When the user asks for a detailed interaction determination, consult rule-consult and label that section as an unofficial rules analysis. Do not make a detailed ruling silently inside a strategy answer.

A live Player 2 move belongs to player2-agent. Deck Coach may supply the deck plan and strategic priorities that Player 2 Agent uses, but does not own the session ledger or confirm an action's legality.

Closed-loop workflow

For a reproducible review, use ${CLAUDE_SKILL_DIR}/scripts/deck_coach_pipeline.py. Its minimum input is decklist + environment + format + player_level + Legend; the input schema is ${CLAUDE_SKILL_DIR}/schemas/deck-coach-input.schema.json.

python ${CLAUDE_SKILL_DIR}/scripts/deck_coach_pipeline.py run `
  --case-id DC-RNG-GLOBAL-001 `
  --output-dir deck-coach-output

That one command writes normalized input, deck-profile.v1, recommendation-mask.v1, an eight-section baseline primer, and deck-coach-evaluation.v1.

Structured observation

The profile computes card-name lookup coverage (what fraction of the decklist's copies matched a card in the bundled snapshot by name -- a data-resolution measure, unrelated to rules-engine coverage), environment/region/set pool, curve, printed Domains and Power costs, Unit/Spell/Gear mix, Battlefield package, interaction/draw-recursion/movement density, eight-role distribution, engine candidates, data provenance, and confidence. Role and engine detection are labelled text heuristics; do not promote them to verified gameplay claims without evidence.

When an R3 behavior manifest is supplied, read ${CLAUDE_SKILL_DIR}/references/shared/card-behavior-coverage.md. The profile adds bounded Main Deck clause coverage, but it must remain separate from card lookup confidence and strategic evidence. With no compatible active manifest, the correct status is unavailable, not an inferred coverage percentage.

Recommendation mask

Before naming a replacement, pass it through the mask. It excludes candidates that are unreleased in the target environment, banned in the selected format, outside the Legend's Domain identity, absent from a supplied collection, supported only by a different source environment, or paired with stale pre-errata text. The registry is dated and still requires a live official check for a real event; provisionally_clear is not an official deck-registration ruling.

Evaluation and primer battle

${CLAUDE_SKILL_DIR}/data/deck_coach_cases.json contains executable cases with input deck, environment, player level, required engine/weakness recognition, forbidden claims, acceptable uncertainty, expert reference primer, and dated sources. suite runs all cases; evaluate scores another candidate; battle compares two versioned/model-labelled candidates using blind A/B labels and accepts a separately recorded expert preference.

The deterministic grader covers the seven declared dimensions and produces useful regression metrics, but it is a proxy. Expert gameplay quality must still be recorded as human scores or blind preference; never call token matching an expert judgment.

deck_coach.py and deck-coach-session.v1 remain available for manually authored reviews. They are artifact editors, not the closed-loop profiler.

Engine evidence intake (ADR-0006)

A session may carry engine_checks — complete engine-check.v1 envelopes — paired with engine_evidence_scope: rules_consistency_only. Both fields or neither. Deck Coach consumes checks produced elsewhere from explicit structured inputs; it does not build a timing state from primer prose.

python ${CLAUDE_SKILL_DIR}/scripts/deck_coach.py engine-check session.json --check engine-check.json
python ${CLAUDE_SKILL_DIR}/scripts/deck_coach.py engine-check session.json --result timing-result.json --kind timing

Attaching never changes the diagnosis, the primer, or behavior coverage; the runner refuses if it would. A supported check is not evidence that a line is good, an unsupported one means the component abstained, and a decision_required one is shown, never answered here.

Evidence discipline

Keep mechanically derived claims, verified real-play claims, and unverified recommendations visibly separate. Do not manufacture specific matchup, mulligan, or sequencing rules merely to fill a template. Preserve the existing Tier 1/2/3 contract in the underlying books.

Do not calculate, retain, or publish metagame-defining play rates, win rates, or matchup percentages.