What problem does it solve? Installed AI skills stay static even as you learn how you actually use them. This Skill closes the loop by reading your local usage traces and compiling that experience back into sharper skill descriptions, instructions, and few-shot examples. ## Core Features & Use Cases - Trace-Based Scoring: Groups trace records from ~/.stellar-build/traces into spans and labels each span win, neutral, or loss using LLM-as-judge evaluation. - Skill Compilation: Tightens trigger descriptions and instructions, then mines 1-3 winning spans into a regenerable LEARNED block of few-shot examples at the end of each SKILL.md. - Safe and Reversible: Backs up every skill to ~/.stellar-build/backups before editing, writes a user profile to profile.md, logs each run to optimizations.jsonl, and supports undo via stellar-loop restore. - Use Case: After a few weeks of using skills like find-stellar-idea, run the optimizer to refine their triggers and embed examples drawn from your own successful sessions. ## Quick Start Ask the assistant to optimize my skills from usage traces so it scores recent spans and rewrites the qualifying installed skills with learned examples.
