How the AI coach works

A replay file is not a video. It is an event log containing player commands, production, kills, ability usage, positions, and game metadata. The parser converts that log into structured data before any language model reads it.

1. Replay events are parsed into players, teams, build orders, kills, and timelines.
2. Static unit, structure, upgrade, and ability data is added to the event stream.
3. Derived metrics identify battles, power-spike windows, economy gaps, and scouting activity.
4. Build-order matching compares the opening against known archetypes.
5. The model receives a structured summary and returns timestamped coaching notes.

The coach can inspect worker count, supply, collection rate, unspent resources, bases, production events, upgrades, army composition, battle trades, and tactical detections. It does not infer these from screenshots; the values come from the replay event stream and derived timelines.

Replay events are combined with unit and ability reference data: cost, supply, hit points, damage, range, armor tags, cooldowns, energy costs, and caster relationships. This lets the coach evaluate army value, missing counters, and whether a spellcaster or tech transition was available before a fight.

Tactical labels are detected before model generation by deterministic rules. Examples include extractor tricks, cannon rushes, drops, recall escapes, turtling, fast expansions, power-droning, proxies, and one-base all-ins. These labels prevent common false positives, such as treating an extractor trick worker dip as a macro error.

The build matcher compares each player's opening against a catalog of known archetypes. A build name is used only when the match confidence is high. Otherwise, the coach describes the observed shape: expansion timing, gas count, production count, tech branch, and first pressure window.

• Frame-perfect micro grading is limited; the replay data records events and sampled positions, not a full coaching VOD.
• Co-op replay visualization is supported, but ladder coaching assumptions do not fully apply to commander-specific units and mission objectives.
• Off-meta openings may match weakly against the build catalog and are described by observed structure rather than by name.
• Team-game MMR data is partial when the replay file does not include all player ratings.