dota-domain
Dota 2 strategic domain knowledge for this project. Use whenever working with heroes, roles, positions, lanes, items, facets, talents, damage types, crowd control, power spikes, tempo, farm priority,
SKILL.md.claude/skills/dota-domain/SKILL.md
GitHubDota 2 Domain Knowledge
Apply this skill whenever a task touches Dota 2 strategy or game concepts — designing a data model, writing a recommendation heuristic, choosing ML features/labels, or reviewing whether an output is strategically sound.
Core principle: context beats aggregate stats
A hero or item's global win rate is a weak, often misleading signal in isolation. The same hero can be excellent in one draft and a liability in another. Any recommendation-shaped decision (hero pick, item build, timing advice) must be evaluated against the actual situation, not a single number.
Never recommend a hero solely from global win rate. Never recommend an item solely from item win rate.
Instead, ground recommendations in as many of these as are available:
- the hero and their kit (see
references/heroes-and-roles.md) - role/position being played (pos 1-5 carry different priorities)
- current patch (balance changes shift what's good)
- skill bracket (a herald-tier read differs from an immortal-tier read)
- player comfort/experience with the hero — a worse hero played well often beats a stronger hero played poorly
- game minute / game state (early laning vs. late-game teamfight phase)
- the draft — both teams' picks, not just your own
- current items already owned and gold available
- observed threats — specific enemy heroes, their items, and their capability (e.g. "do they have a burst combo that kills me from full HP?")
If the surrounding code or a caller wants a recommendation, check whether enough context is present to reason about it this way. If it isn't, that's a missing-input problem to raise, not something to paper over with a global average.
Separate mechanics from strategy
Mechanics = objective, ruleset facts: what an item does, a hero's base stats, how buyback cost scales, whether a spell pierces spell immunity.
Strategy = situational judgment built on top of mechanics: whether buying that item now is correct, whether a fight is winnable, whether to push high ground.
Keep these separate in code and in reasoning:
- Mechanical facts belong in structured data (see
references/data-model.md), not hardcoded into strategic logic. - Strategic logic consumes mechanical facts plus game state; it should not need to re-derive or guess at mechanics.
This mirrors the GSI layering in this project (see the dota-gsi skill):
raw state flows up into a normalized MatchState; strategic reasoning is a
separate layer built on top of it, never mixed into ingestion/normalization.
Prefer structured domain concepts over free text
When representing hero roles, damage types, CC types, item timings, draft
state, etc. in code, ML features, or prompts, use structured, enumerable
concepts (e.g. a Position of 1-5, a DamageType enum, a CrowdControlType
enum) rather than ad hoc free-text strings. Structured concepts are testable,
composable, and prevent subtle inconsistency (e.g. "carry" vs "hard carry" vs
"pos 1" meaning the same thing three different ways in three places).
Reference files
Load these as needed — don't pull them into context for tasks that don't need the detail:
references/heroes-and-roles.md— roles, positions, lanes, hero archetypes, damage types, crowd control taxonomyreferences/game-flow.md— power spikes, tempo, farm priority, item timings, Roshan, high ground, buyback, team fights, win conditionsreferences/draft-and-matchups.md— draft phases, counters, synergies, matchup reasoning, facets and talentsreferences/data-model.md— suggested structured representations for these concepts in this project's domain layer, and how they relate to the GSIMatchState/DomainEventspipeline