Skill Dispatch Guide

When a hockey analytics request comes in, match it to a pattern below. Load those skills. Don't load everything.

Default data tool: puckapi-tool -- covers most data needs. Load it alongside the relevant strategy skill unless the request specifically needs a different tool.

First-Use Persona Detection

On first interaction, identify which path the user is on:

Request Patterns -> Skills to Load

"Who plays tonight?" / "What's the schedule?" / "Game results"

1. game-lookup -- find games by date, team, season
2. puckapi-tool -- get_games, get_schedule endpoints

"How are the Sabres doing?" / "Standings" / "Team stats"

1. team-analysis -- standings, stats, rankings, SOS
2. puckapi-tool -- get_standings, get_team_stats

"Tell me about Connor McDavid" / "Player stats" / "Compare players"

1. player-scouting -- search, stats, comparison, NHLe
2. puckapi-tool -- search_players, get_player_stats

"Who's starting in goal?" / "Goalie matchup" / "Save percentage"

1. goalie-analysis -- leaderboard, workload, xG-adjusted metrics
2. puckapi-tool -- get_goalie_stats

"What are the odds?" / "Line shopping" / "Best price"

1. odds-explorer -- multi-book comparison, line movement
2. puckapi-tool -- get_odds, get_line_movement
3. For odds math/devigging: also load odds-analysis

"Show me games where..." / "Find all teams that..." / natural language query

1. nl-to-query -- translate natural language to structured queries
2. puckapi-tool -- routes to appropriate endpoint
3. If pattern found: ai-hockey-workflow for follow-up hypothesis testing

"What is Corsi?" / "Explain xG" / "Hockey analytics basics"

1. hockey-analytics -- metric definitions, formulas, context
2. puckapi-tool -- pull real data to demonstrate

"How do I use Claude for sports analysis?" / "What questions should I ask?"

1. ai-hockey-workflow -- prompt patterns, hypothesis testing, iteration
2. Load relevant data skill based on the specific question

"How do I automate this?" / "Run daily" / "Pipeline"

1. data-pipeline -- GitHub Actions, scheduling, model versioning
2. Reference puckapi-tool for endpoint costs in automation budget

"Help me build features" / "Feature engineering" / "What features should I use?"

1. feature-engineering -- rolling windows, leakage detection, shift(1)
2. puckapi-tool -- historical data endpoints
3. If user is new to metrics: also load hockey-analytics

"How do I validate my model?" / "Is k-fold okay?" / "Cross-validation"

1. walk-forward-validation -- temporal CV, anti-leakage, significance testing
2. If user has a model: probability-calibration for probability verification

"Help me build a prediction model" / "Train a model" / "XGBoost"

1. model-building -- model selection, training, evaluation
2. walk-forward-validation -- correct evaluation method
3. feature-engineering -- if features aren't built yet

"Build an Elo rating system" / "Rating system" / "Team ratings"

1. elo-engineering -- 5 variants, tuning, carryover
2. puckapi-tool -- get_games for historical results

"Are my probabilities calibrated?" / "Platt scaling" / "Brier score"

1. probability-calibration -- reliability diagrams, scaling methods
2. If model outputs available: proceed directly
3. If not: route to model-building first

"How do I devig odds?" / "Implied probability" / "Shin method"

1. odds-analysis -- conversion, devigging, no-vig lines
2. puckapi-tool -- get_odds for real numbers to work with

"Should I bet this game?" / "Where's the edge?" / "Expected value"

1. edge-detection -- EV calculation, Kelly sizing, CLV
2. odds-explorer -- current market odds
3. probability-calibration -- verify model probabilities first
4. Requires: user must have a calibrated model

"Backtest my model" / "Historical performance" / "Is my edge real?"

1. backtesting -- walk-forward backtest, strategy simulation
2. puckapi-tool -- historical odds (heavy credits)
3. If model health check: backtesting includes model audit

"Build an xG model" / "Expected goals from scratch" / "Shot quality model"

1. xg-model-building -- full xG pipeline from play-by-play
2. If user needs basics first: hockey-analytics for xG concept

"Over/under prediction" / "Totals model" / "Will this game go over?"

1. totals-modeling -- pace, goalie matchup quality, under bias
2. puckapi-tool -- get_odds for totals lines

"Player props" / "Will McDavid score?" / "SOG prop" / "DFS"

1. prop-modeling -- TOI projection, per-60 rates, matchup adjustment
2. player-scouting -- player data
3. puckapi-tool -- get_player_stats

"Build WAR" / "Player value" / "Contract analysis" / "RAPM"

1. war-gar-decomposition -- RAPM ridge regression, component GAR
2. player-scouting -- player data for context

"Playoff odds" / "Will they make the playoffs?" / "Season simulation"

1. playoff-simulation -- Monte Carlo, bracket simulation
2. elo-engineering -- if rating system needed as input
3. puckapi-tool -- remaining schedule

"Preview tonight's game" / "Sabres vs Leafs preview"

1. game-preview -- compound skill, loads data internally
2. puckapi-tool -- multiple endpoints (~34 credits)

"Tonight's card" / "Full slate" / "All games tonight"

1. daily-card -- slate analysis, edge rankings
2. Requires: user's model for edge calculations
3. Heavy credits: ~10 credits per game on slate

"Track my bets" / "Am I profitable?" / "CLV" / "Log this bet"

1. bet-tracker -- prediction logging, CLV, significance
2. If performance declining: route to backtesting for audit

"Make a chart" / "Visualize" / "Calibration plot" / "Player card"

1. visualization -- chart type selection, code generation
2. Requires output from another skill (calibration, backtest, preview, etc.)

When to Load Docs

• Unknown hockey term -> docs/hockey-glossary.md
• Unknown betting term -> docs/betting-glossary.md
• Need to pick a data tool -> docs/tool-routing.md
• Date/season confusion -> docs/season-logic.md

When NOT to Use This Guide

Single-skill requests don't need dispatch. If someone says "devig these odds: +150 / -180" -> just load odds-analysis.