@chad and I dive into advanced Viking landing logic, edge-case micro, and how to avoid idle air supply. They also break down Drekken’s experiment building a ladder-ready bot using vibe-coding with Claude, Grok, Gemini, and GPT-5, comparing accuracy, speed, error rates, and framework usage.
Key Takeaways:
Dynamic Viking micro: land when no air or Colossus threats are present, lift when air appears, and use special handling for tanks, splash zones, and terrain height mismatches.
Avoid “transform traps”: Vikings take damage while landing, so proximity to anti-air structures or ranged threats must be part of the landing heuristic.
Use demand-based counters to scale Viking production, weighted by enemy air types, rather than static ratios.
Vibe-coding bots: Claude produced the most reliable code with the fewest errors, used search intelligently, and followed framework rules; Grok was fast but sloppy; Gemini improvised solutions but ignored frameworks; GPT-5 Codex was slow but methodical.
Micro-ladder testing revealed strengths and weaknesses across models in stutter-step, target selection, and handling edge cases like Hellion micro and Reaper jumps.
AI coding improves dramatically when you supply a structured plan, global rules, framework references, and example bots to anchor the model.
Leashing behavior: Medivac following can work with “unit-passing hooks,” but can leave stranded healers if rally logic or unit flow breaks.