AI coach reference
Tactics catalog
The 25 named plays the engine looks for. Each one is detected by a deterministic rule, not by the AI model — when it fires, the coach reads from this catalog to interpret what the player did and why.
The coaching note on each entry is the most useful field. It tells the model what NOT to flag — the worker dip during a double extractor trick is correct execution, not a macro mistake.
Execution tricks
Small skill plays that look like mistakes if you don't know what they are. The coach is told not to flag the worker dip from a double extractor trick as a macro slip.
Morph two Extractors at the same time while approaching the 14/14 supply cap, then cancel one for a 75% mineral refund. The temporary Drone reservation lets you avoid being supply-blocked for a single tick at the supply transition without sacrificing the Drone permanently.
- Why a player does it
- Sneak in one extra Drone or Zergling before your Overlord pops, smoothing the early supply curve.
- What it costs
- A few seconds of mineral lockup and one cancelled Extractor's worth of Drone time. Trivial cost when timed correctly.
How it's detected
Two signals, either fires the tactic. Primary: an Extractor in the player's base tracks built before t=180s with destroyedAt-builtAt between 3 and 25 seconds (the cancelled half of the trick). Secondary fallback for replays without lifecycle data: two Extractors in the build order with start times ≤3 seconds apart, both before t=120s. Zerg only.
Sitting at max supply with spare minerals for an extended window — either the player is supply-blocked (mistake) or deliberately faking poverty / saving for a tech transition (intentional).
- Why a player does it
- When intentional: bait the opponent into a misread of your economy, save up for a tech swap (e.g. mass remax), or hold position before a planned engagement.
- What it costs
- Wasted production cycles. Every second floated is a unit not built. Only worthwhile when the saved bank converts to a decisive remax or tech jump.
How it's detected
Walks the timeline once per player. Tracks any continuous run where supply == supplyMax AND mineralsUnspent > 400. Any run lasting ≥60 seconds fires the tactic.
Macro and tempo
Economy and tempo decisions. Power droning, fast expansions, turtling. The shape of how a player chose to spend the early game.
A second base placed unusually early — at or under the race-specific supply benchmark (Terran ~17, Protoss ~21, Zerg pool-before-hatch ~14, or hatch-first ~12 supply). Defines the player's economy-first commitment.
- Why a player does it
- Trade safety for income. A fast natural roughly doubles mining within 1-2 minutes; the lead compounds through the mid game if it survives the early-game pressure window.
- What it costs
- Vulnerability to early all-ins, cheese, and proxy plays. Requires good scouting to confirm the opponent isn't punishing the greed.
How it's detected
Reads the player's town halls in order of completion. The second hall must exist, and at the first timeline snapshot at or after its completion the player's supply has to be ≤ the race's supplyMax: 17 for Terran (CC/OC), 19 for Protoss (Nexus), 16 for Zerg (Hatchery).
Heavy Drone production through the early game — typically 60-70% of all units made are Drones in the first 4 minutes. The defining shape of an economy-first Zerg opener.
- Why a player does it
- Maximize mining capacity early so the Hive-tier transition can come at advantageous timings. Trades off early army value for compounding economic advantage.
- What it costs
- If aggression hits before army units come online, a power-droning Zerg has minimal defense and can fold. Requires a confirmed read on opponent NOT cheesing.
How it's detected
Walks the build order until t=240s. Counts Drones vs army units (Zergling, Baneling, Roach, Ravager, Hydralisk, Lurker, Mutalisk, Corruptor, BroodLord, Ultralisk, Infestor, Viper, SwarmHost). Fires when total drones+army ≥15, army ≤4, AND drones/(drones+army) ≥0.85. Zerg only.
Defense
Defensive postures. The coach treats heavy static defense as a deliberate read on the opponent, not a default.
Heavy investment in static defense (10+ structures: bunkers, photon cannons, spine/spore crawlers, missile turrets) past the 10-minute mark, with army value held below ~4000 through the first 15 minutes.
- Why a player does it
- Survive without committing army. Delay the engagement until tech (Carriers, Tempests, BC, Lurker, broodlord, etc.) renders the opponent's earlier composition obsolete.
- What it costs
- Map control surrendered. Production efficiency suffers (static defense is a pure mineral sink with no mobility). Loses if opponent expands faster than you tech.
How it's detected
Only checks games of durationSeconds ≥1080. Counts the player's static defense (PhotonCannon, SporeCrawler, SpineCrawler, Bunker, MissileTurret, PlanetaryFortress, ShieldBattery) where destroyedAt is null OR > 600s — i.e. still standing past the 10-minute mark. Needs ≥6 such defenses. Then walks the timeline up to t=900s and fires only when armyValue exceeded 4000 in less than 40% of those snapshots (≥5 snapshots required).
All-ins
Committed mid-game pushes. If the all-in lands, it wins. If it bounces, the player is usually behind for the rest of the game.
Committed push where worker production stalls, mineral bank approaches zero, and army value is held at or above 75% of the player's peak for 60+ seconds while engaging the opponent. The player is betting the game on this engagement.
- Why a player does it
- Win or lose the game in a single committed push. Used when the player believes their current army+composition power-spike is decisive — usually because of an opponent timing read or a scout-confirmed tech window.
- What it costs
- If the all-in fails, the player has near-zero economy to rebuild from and the opponent has free run. There is no 'safe lose' from an all-in.
How it's detected
Six-snapshot sliding window (~60 seconds). Skips players whose peak armyValue across the game stayed under 1500. Within the window: workerCount delta ≤ +1 (no real worker production), workerCount at window start ≥20 (skips the very early game), mineralsUnspent < 300 in at least 70% of the seven snapshots, armyValue ≥ 0.75 × peakArmy in at least 70% of them, AND opponent unitsLost summed across the window plus 30 seconds ≥5. Fires at most once per player.
Mining rate drops sharply while worker count stays roughly flat — workers redirected to defend an attack or join an offensive push. Distinct from worker losses (where the count would also drop).
- Why a player does it
- Defensive: turn workers into emergency combat units when an attack would otherwise break through. Offensive: stack damage on a desperate push when the army alone can't finish.
- What it costs
- Mining stops while the workers are out. Several pulled SCVs / Probes / Drones die in most fights. Recovery time depends on whether the pull saved the game or wasted the workforce.
How it's detected
Compares each timeline snapshot to the snapshot two ticks earlier (~20 seconds). Pull signature: mineralCollectionRate dropped by ≥35% AND workerCount dropped by ≤2 (so workers didn't actually die, they just stopped mining). Skips windows where prior workerCount <14 or prior mineralCollectionRate <400. Throttled to one fire per player every 60 seconds. A second pass reclassifies generic pulls as defensive harass response, base relocation, or real worker rush when context lines up.
Cheese
All-or-nothing early aggression. Cannon rushes, proxies, gas steals — bets the entire game on the opponent not having the right scout.
Forge + Photon Cannons placed inside or right next to the opponent's main base, often before the opponent's first defensive structures or units are online.
- Why a player does it
- End the game in the first 4-6 minutes by walling the opponent into their main with cannons. Even a partial success damages the opponent's economy enough to fall behind permanently.
- What it costs
- If the cannon rush fails, the Protoss player has spent ~500 minerals on a failed all-in with no economy or tech to fall back on. Almost always loses.
How it's detected
Protoss only. For each Forge or PhotonCannon built before t=300s with distance from the player's first Nexus ≥0.35 × map diagonal, treat as a forward candidate. Forwards are clustered: same-cluster requires Euclidean distance ≤15 tiles AND at most 30 seconds between the new candidate and the cluster's most recent member. Each cluster fires one event regardless of how many cannons it contains.
A town hall — Hatchery, Command Center, or Nexus — built closer to an opponent's main than to the player's own, in the opening. Sometimes true cheese (proxy roach pressure, proxy planetary, hidden Nexus into proxy gates), sometimes a high-risk fast-expansion that lands on the wrong half of the map. Either way it's a town hall placed for aggression, not a clean third-base macro choice.
- Why a player does it
- Forward production puts units on the opponent's doorstep without the cross-map walk. The aggressor trades safety (no defensive structures at home) for early pressure that the opponent's first defensive units aren't yet positioned to handle.
- What it costs
- If scouted and killed, the player has spent ~300-400 minerals on a building that produces nothing. If the follow-up aggression doesn't pay off, the player is also missing the natural expansion they would have taken with that money — they fall behind on economy without the trade.
How it's detected
Race-agnostic. For each Hatchery / CommandCenter / Nexus the player built, finds the nearest opponent main and the player's own first main. Fires when the distance to the nearest opponent main is less than the distance to the player's own main. No fixed map-size threshold.
Production buildings (Barracks, Factory, Gateway, Stargate) built far from home — typically near or behind the opponent's base — to shorten the reinforcement path for an early all-in.
- Why a player does it
- Hit timing windows that wouldn't be possible from a normal main-base build (proxy 2-rax marines, proxy Stargate Oracle, proxy Robo Immortal pressure).
- What it costs
- The proxy buildings are easy to scout-and-kill. If detected and dealt with cleanly, the player has lost time and minerals with nothing to show for it.
How it's detected
For each production building (Barracks, Factory, Starport, Gateway, Stargate, RoboticsFacility, Hatchery, SpawningPool, RoachWarren) built before t=240s, computes Euclidean distance from the player's first main. Fires when that distance ≥0.45 × map diagonal. Town-hall proxies are handled by proxy_townhall instead.
An Assimilator / Refinery / Extractor placed on one of the opponent's vespene geysers, denying them gas income.
- Why a player does it
- Slow the opponent's tech transition by 30-90 seconds. Used as a soft cheese that doesn't fully commit but disrupts a known tech-dependent build.
- What it costs
- Costs 75 minerals; gives away the strategic intent (opponent now knows you're playing aggression). The opponent can also kill the gas steal building, refunding nothing.
How it's detected
Checks Refinery / Assimilator / Extractor structures with builtAt ≤180s. Fires when the geyser is ≥25 tiles from the player's own first main AND within 20 tiles of any opponent's first main. Both conditions are required.
Every worker pulled and sent to attack the opponent's mineral line in the first 2-3 minutes. Mining collapses on the attacker's side; opponent loses several workers.
- Why a player does it
- Immediate game-ending all-in. If the worker rush succeeds in killing all opposing workers, the opponent cannot recover; if it fails, the attacker has no workers and loses to any normal opener.
- What it costs
- Pure all-or-nothing — there's no recovery from a failed worker rush.
How it's detected
Establishes a baseline from the first timeline snapshot in the 30–60 second window: that snapshot's mineralCollectionRate must be ≥500. Then between t=60s and t=150s, looks for a snapshot where mineralCollectionRate ≤100 AND workerCount ≥6 (workers alive but not mining). Confirms with opponent worker losses (SCV / Probe / Drone) ≥3 inside the snapshot's window − 10s … window + 60s.
Harass
Multi-prong damage attempts. Reaper, banshee, oracle, mutalisk, drop play. The coach reads each as its own named play with its own cost-benefit.
Medivac, Warp Prism, or Nydus Worm delivers a small unit force directly into the opponent's mineral line, far behind their army.
- Why a player does it
- Force a defensive split. Even a failed drop usually trades favorably because the opponent must pull units from their main army to defend, fragmenting their position.
- What it costs
- If the dropped units die without doing damage, the player has wasted ~200-300 minerals of units plus the carrier (Medivac/Prism). Multi-prong play requires 2+ APM-intensive operations simultaneously.
How it's detected
Walks transport tracks (Medivac, WarpPrism, WarpPrismPhasing, Overlord, OverlordTransport). For each, finds the first position where Euclidean distance to any opposing player's first main is <18 tiles. Fires once per transport. We don't track passengers, so a Medivac parked at an enemy main is treated as a drop in 1v1.
One or more Reapers cliff-jumping into the opponent's mineral line — typically 1-2 Reapers in the 2:30-4:00 window. Combat Drugs lets them disengage when threatened.
- Why a player does it
- Pick off a few workers, scout production buildings and tech path, drag one or more defenders away from the army.
- What it costs
- Reapers don't transition well into the mid-game army; the player has invested gas in a unit that becomes obsolete after 5:00. A failed harass means the gas was wasted.
How it's detected
Reaper unit track enters within 20 tiles of any opponent town hall before t=240s. Confirms with either lingering (Reaper still in the 24-tile radius after 3 seconds) OR opponent worker losses within 45 seconds. Pure fly-bys without lingering or kills don't count.
Banshees (with or without Cloak) raid the opponent's mineral line. Typical timing: 5:00-7:00 from a 1-base or 2-base Starport opener.
- Why a player does it
- Worker damage at scale. Cloaked Banshees vs. opponents without detection can wipe entire mineral lines in one pass.
- What it costs
- Banshees are expensive (150/100 each) and weak to anti-air. If detection lands before the Banshee gets value, the gas was misallocated.
How it's detected
Banshee track enters within 20 tiles of an opponent town hall before t=600s. Confirms with ≥4 seconds of lingering (in the 24-tile follow-up radius) OR opponent worker losses within 45 seconds.
Liberator(s) parked over the opponent's mineral line in defender (anti-ground) mode, killing any workers in the targeted zone.
- Why a player does it
- Force the opponent to waste APM evacuating workers, or pay anti-air to kill the Liberator. Every second the Liberator stays alive over their mineral line drains worker count.
- What it costs
- Liberators are slow to position and easy to kill from below if the opponent has even moderate anti-air. Trade is good only if the opponent can't quickly punish.
How it's detected
Liberator (any mode) within 20 tiles of an opponent town hall before t=720s. Confirms with ≥5 seconds in position OR opponent worker losses within 45 seconds.
Medivac drops one or more burrowed Widow Mines into the opponent's mineral line. Each Widow Mine activation deals massive single-target + splash damage to workers.
- Why a player does it
- Immediate massive worker damage on first detonation — frequently kills 4-6 workers per mine before detection. Lower micro overhead than a Banshee harass.
- What it costs
- If the opponent has detection or the Mines are scouted before burrow, the Widow Mines die without value. The Medivac is also exposed during the drop.
How it's detected
Widow Mine appearing on the map (the in-Medivac form has no position; only the dropped form does) within 20 tiles of an opponent town hall before t=720s. Confirms with ≥2 seconds in place OR opponent worker losses within 30 seconds.
2-4 Hellions running into the opponent's mineral line — typically against Zerg / Protoss workers (Light armor). Hellions deal +6 vs Light, melting workers.
- Why a player does it
- Worker damage and forcing defensive structures. Scout the opponent's gas and tech path while doing it.
- What it costs
- Hellions are fragile and lose to massed Roaches / Stalkers. If they get caught or if the opponent's anti-light response is ready, the Hellions die without value.
How it's detected
Hellion / Hellbat (HellionTank / BattleHellion) within 20 tiles of an opponent town hall before t=360s. Confirms with ≥3 seconds in place OR opponent worker losses within 45 seconds.
Multiple Adepts using Psionic Transfer (the shade) to skip past defenses into the opponent's mineral line. Adepts deal +12 vs Light, lethal to workers.
- Why a player does it
- Worker damage + scouting + drag defenders out of position. Shade can also be used purely for vision without committing the body.
- What it costs
- Adepts are weak in direct fights and lose efficiency once the opponent has detection / strong anti-light units (Hellbat, Marauder, Roach).
How it's detected
Adept (or its AdeptPhaseShift projection) within 20 tiles of an opponent town hall before t=600s. Confirms with ≥3 seconds in place OR opponent worker losses within 45 seconds.
Oracle(s) flying into the mineral line in Pulsar Beam mode. Oracles deal +7 vs Light per shot, killing a worker every 2 shots.
- Why a player does it
- Heavy worker damage in the 4-6 min window before the opponent can have detection up. Stasis Wards (later upgrade) add area denial.
- What it costs
- Oracles are gas-heavy (150/150) and become economic dead weight if they don't get value. Detection or queens / spore crawlers shut them down hard.
How it's detected
Oracle within 20 tiles of an opponent town hall before t=420s. Confirms with ≥3 seconds in place OR opponent worker losses within 45 seconds.
Cloaked Dark Templar arrive in the opponent's main or natural before detection is in place. DT autoattack of 45 one-shots most workers.
- Why a player does it
- Game-ending worker damage if undetected — easily 10+ workers per DT before they die. Even when detected, the defender has lost time and resources to detection / response.
- What it costs
- If detection (Observer / Overseer / Missile Turret / Photon Cannon) is up before the DT arrives, this trades terribly — DTs are 125/125 and brittle to direct fights.
How it's detected
Dark Templar within 20 tiles of an opponent town hall before t=600s. Confirms with ≥3 seconds in place OR opponent worker losses within 45 seconds. The catch is the opponent has no detection — a single DT often kills several workers before being spotted.
A Mutalisk flock (5+) flying around the map killing isolated workers, defenders, and supply structures. Their Glave Wurm bounces between targets.
- Why a player does it
- Long-term economic and tempo pressure. Mutalisks regenerate health and can disengage on any side of the map at high speed; the opponent must defend everywhere simultaneously.
- What it costs
- Mutalisk flock is very gas-intensive (100 gas each). If the Mutalisks die in a single bad engagement, the Zerg is significantly behind. They're also weak to focused anti-air (Phoenix, Thor HIP, Marines + Medivac).
How it's detected
Mutalisk within 20 tiles of an opponent town hall before t=720s. Confirms with ≥4 seconds in place OR opponent worker losses within 45 seconds.
Lurkers burrowed near the opponent's mineral line or expansion. Line-AOE attack devastates clumped workers and unupgraded armies.
- Why a player does it
- Area denial with unkillable units (without detection). Forces the opponent to stop moving units through certain zones and to commit to detection.
- What it costs
- Lurkers are immobile while burrowed. Detection + ranged units (Vikings, Stalkers, Tanks) deletes them quickly. Cost is high — 50/100 + 0 supply. The investment is binary (works or doesn't).
How it's detected
Lurker (or its LurkerMP / LurkerMPBurrowed forms) within 20 tiles of an opponent town hall before t=720s. Confirms with ≥5 seconds in place OR opponent worker losses within 60 seconds.
A pack of Zerglings (typically 8-16) bypasses the opponent's army by running around the map edge and arriving at the natural / third base mineral line.
- Why a player does it
- Worker damage on a base the opponent hasn't fully defended. Forces a defensive split and trades minerals for minerals favorably (a Zergling is 25 minerals, a worker is 50/75).
- What it costs
- Most of the runby Zerglings die to defensive structures or pulled defenders. Only worth it if 4+ workers die.
How it's detected
Zergling unit track within 20 tiles of an opponent town hall before t=300s. Confirms with ≥3 seconds in place OR opponent worker losses within 30 seconds.
Transitions
Movement plays. Recall escapes, repositioning before a fight.
A group of units suddenly arrives at a cluster point — Mothership Strategic Recall, Nexus Strategic Recall, or Nydus Worm pop. Detected by a tight unit cluster appearing without a continuous movement trail.
- Why a player does it
- Surprise reinforcement at a different location, evacuation from a losing fight, or split-second reposition before a sieged engagement.
- What it costs
- Strategic Recall has cooldowns and energy costs; Nydus Worm requires the canal to live until units exit. A failed Nydus (canal killed mid-emerge) loses both the worm and the units inside.
How it's detected
Walks every non-worker unit track. Looks for jumps of ≥80 tiles between consecutive position samples taken within 2 seconds of each other. Then groups jumps with the same pid, within a 2-second window of the seed jump's t, whose destinations land within 12 tiles of the seed's destination. Fires when at least 4 units land in the same cluster — that's the giveaway for Recall vs ordinary fast army movement.