A drop-in skill bundle for loss-function-driven (LFD) agentic loops — turn a high-level goal into a paste-able /goal prompt, scaffold a harness, and run the outer loop until the inner agent's skill converges.

Small local models doing real agentic coding: pi + Ollama + an LFD harness that moves the intelligence out of the model and into the loop. Build log included.