@chad walked through his SCV repair micro for defending worker rushes, @mindme debuted his new Zerg bot Tito and showed how early expansion blocking snowballs an economic lead, and I demoed my autonomous Ralph Loop setup for letting an LLM iterate on bot code overnight. The session also sparked a meaty debate on how much build orders really matter.
Key Takeaways
- Repair micro beats raw HP in worker fights. DogTato’s SCVs outlast Probes by rotating injured units to repair while healthy ones stay in front. Keep a minimum number of workers mining so you don’t tank your economy even if you win the trade.
- Mineral walking is the Protoss answer. MindMe swaps low-health Probes out through mineral walk, cycling fresh workers in. The 50% HP threshold is a solid default for both Terran repair and Protoss retreat decisions.
- Expansion blocking with Zerglings compounds fast. MindMe’s burrowed Zerglings on third and fourth bases delayed Eris long enough to build a 20-worker lead that snowballed into a win, even without strong micro.
- Build orders are the foundation, not the finish line. You may not finish the build, but the first few minutes set up exponential advantages. MindMe’s Negative Zero succeeded because of disciplined openers, not flashy late-game code.
- Scouting gaps lose games more than bad compositions. Negative Zero lost to Avocados largely because it didn’t see the early expand and triggered an unnecessary all-in response. Better scouting equals better build order selection.
- The Ralph Loop can automate chunky bot features. Agent MD + a scoped plan + a while loop that restarts a fresh LLM context each iteration. Pair it with a JSON event logger and a validation script so the loop knows when to stop.
- LLM-driven strategy is coming. The TextStarCraft2 environment translates game state to text and LLM commands back to actions. Running it async on a small local model for strategic decisions (not micro) looks feasible and cheap.