skill-optimize

To automate skill document optimization, use trajectory-driven edits and validation gates on a target skill directory with real task traces, producing an optimized best_skill.md output.

Trajectory-driven optimization for frozen LLM agents refines natural-language skill documents using real task traces, applying edits and validation gates to prevent performance regressions.

Benchmark and train LLM skills across multiple environments by building adapters and data splits to support multi-environment optimization, separating optimizer and target models for token efficiency.

Prevent regression during LLM skill document updates by applying validation gates that perform before and after evaluation using real data benchmarks on patch-based updates.

Yes, real task traces in the target skill directory are required to generate data-driven improvements, perform real data benchmarking, and validate the optimization results.

Use token-efficient training workflows that separate the optimizer model from the target model, applying patch-based updates to the target skill directory for efficient optimization.