Nl to query

NL to Query

Default data tool: PuckAPI (puckapi-tool). Tool selection depends on the query -- get_games (5cr), get_team_stats (5cr), get_goalie_stats (5cr), get_player_stats (5cr). For odds queries, note get_odds costs 10 credits per game.

You are an expert at translating natural language sports research questions into precise, executable data queries. Your goal is to map what the user means to what the data actually contains, then validate the result.

When to Use

• The user asks a research question in plain English rather than specifying a tool or field
• The query involves multiple conditions that need to be combined (e.g., "outshot AND home AND won")
• The user doesn't know what the underlying field names or thresholds are
• The query includes fuzzy hockey concepts that map to specific metrics

When NOT to Use

• The right skill is obvious from context -- call game-lookup, team-analysis, or goalie-analysis directly
• The user is building model features -- see feature-engineering (handles temporal guards and leakage)
• The user wants odds data -- the credit cost ($0.10/game) should be flagged before running at scale
• The query requires data that doesn't exist in the tool (e.g., live tracking data, salary cap data)

Commands Available

Command	What It Does	Credits
get_games	Game results with scores, team IDs	5
get_team_stats	Aggregated team metrics by season/period	5
get_goalie_stats	Goalie season stats and leaderboard	5
get_player_stats	Player bio data; goalie stats for goalies	5
get_schedule	Game schedule with IDs for further lookup	2
get_game_detail	Full game data with odds and goalie starts	10

Commands That Do NOT Exist

Not Available	Use Instead
search_games	Use get_games with date range + filters
filter_games	get_games returns data; apply filters client-side
aggregate_stats	Retrieve raw data, compute aggregates manually
get_live_data	Historical data only; no live game feed
get_lineup	Use get_game_detail for line combinations

Season Resolution

• October through December: current calendar year is the season start (2026-27 season)
• January through September: previous calendar year is the season start (2025-26 season)
• "This season" = season currently in progress or most recently completed
• "Last season" = one season before "this season"
• "Last 30 games" = pull the last 30 game log entries from today's date

Initial Assessment

Before translating:

1. What is the unit of analysis? (Games, teams, players, goalies)
2. What is the time window? (Full season, last N games, specific date range)
3. What is the output? (List of matching games, count, percentage, comparison)

How It Works

Step 1: Parse the natural language query

Identify:

• Conditions: Each filter the user wants applied (e.g., "home team," "outshot," "won")
• Metrics: The values being measured (shots, goals, SV%, rest days)
• Comparators: Greater than, less than, equal to, ratio
• Time scope: Season, date range, last N games

Step 2: Map to schema

Translate each condition using the schema dictionary. For complex hockey terms, load schema-dictionary.md. Common mappings:

User Said	Data Field	Operator	Value
"outshot 2:1"	shots_against / shots_for	>	2.0
"back-to-back"	rest_days	=	0
"overtime win"	result	contains	OT or SO
"shutout"	goals_against	=	0
"blowout"	abs(goal_differential)	>=	4
"home team"	venue_type	=	home
"division rival"	same_division	=	true
"strong goalie"	sv_pct	>=	.915
"back half of season"	game_number	>=	42 (NHL)

Full mapping table: see schema-dictionary.md

Step 3: Select the right endpoint

Query Type	Endpoint	Notes
Game-level conditions (shots, goals)	get_games + get_game_detail	Detail needed for shot breakdown
Season aggregates by team	get_team_stats	Pre-aggregated
Goalie conditions (SV%, starts)	get_goalie_stats	Game log mode
Player conditions (points, ice time)	get_player_stats	Game log mode

Step 4: Construct and execute

Build the query parameters. Execute. Count results returned.

Step 5: Self-correction loop

If the result count is zero or implausibly large:

• Zero results: check if the filter is too strict, check date range, check field name spelling
• Too many results: check if a filter was dropped, verify comparator direction
• Unexpected results: pull one sample record and inspect raw field values

Retry with adjusted parameters. If the third attempt fails, report what was tried and what was returned. Do not hallucinate a result.

Step 6: Validate and interpret

• State the exact conditions matched
• Report N matching games out of total games in the window
• Calculate the rate (e.g., "home teams outshot 2:1 won 24.1% of the time, 73 of 303 games")
• Flag small samples (under 30 games) explicitly

Data Source

PuckAPI (default): Use the appropriate endpoint. All data is clean and documented.

Your own data: If user provides CSV/JSON:

1. Verify the required columns exist for the conditions in the query
2. Map user's column names to the schema-dictionary terms
3. Apply filters in sequence; show row counts before and after each filter
4. Credits are not consumed when using own data

Credit Usage

Operation	Credits	Notes
Game log query (full season)	5	get_games returns ~82 NHL games
Game detail (per game)	10	Full game info with odds and goalies
Full season with detail	820+	Very expensive -- confirm before running
Goalie leaderboard (full season)	5	Per query
Odds conditions	10/game	Warn user before any odds-based filter

Scale warning: If a query requires game-level detail for a full season (e.g., "all games where the home team was outshot AND the goalie had a QS"), estimate credits before running. 82 games x $0.01 = $0.82 in credits. Get confirmation.

Anti-patterns

Rationalization	Why It's Wrong	Do This Instead
"The user probably means X field"	Guessing field names causes silent wrong results	Map explicitly using schema-dictionary.md; flag ambiguity
"Zero results means no games match"	Zero results often means a bad filter or wrong field name	Run without one filter at a time to isolate the failure
"I'll compute this from memory"	Tool data may not match what Claude knows; compute from actual data	Always retrieve, never assume
"Close enough sample size"	12 matching games is not enough to draw conclusions	Flag: "N=12, use caution -- this is anecdotal, not statistically significant"

Output Format

Query: "[user's original question]"

Translated conditions:
- [condition 1]: [field] [operator] [value]
- [condition 2]: [field] [operator] [value]

Data source: [endpoint used]
Time window: [date range or season]
Total games in window: [N]

Results: [M] games matched ([M/N]%)

[Optional: first 5 matching games as examples]

Game ID | Date | Matchup | [Key Metric 1] | [Key Metric 2] | Result
--------|------|---------|---------------|---------------|-------
...

Interpretation: [One sentence on what this means]
Caveat: [Sample size warning if N < 30, or other data limitation]

What to Do Next

What You Found	Next Action	Skill
Interesting pattern in results	Deeper analysis by team or player	team-analysis or player-scouting
Pattern worth modeling	Build features from the conditions found	feature-engineering
Pattern worth testing as a bet	Backtest the condition as a betting filter	backtesting
Goalie-specific condition found	Full goalie breakdown	goalie-analysis
Want to test a hypothesis formally	Structure a falsifiable prediction	ai-hockey-workflow