5 GitHub repos AI engineers don't talk about but should. stanfordnlp/dspy : 34K ⭐ You don't write prompts. You write code that writes prompts. Game over for manual tuning. jxnl/instructor : 6M downloads/month Structured outputs for every LLM. Every serious builder uses this quietly. eth-sri/lmql : ETH Zurich Constrain model output at generation time. Not post-filtering. Actual control. repowise-dev/repowise Indexes your codebase with dependency graphs + git history for AI agents. Your agent is currently walking your repo blind. This fixes that. walkinglabs/learn-harness-engineering 12 lectures on why capable agents fail in production. $0. No upsell. Capable ≠ reliable. These repos teach the difference.