Prompt Optimization with Agent Lightning

Use this skill to improve a single markdown prompt file by running Agent Lightning prompt optimization against explicit JSONL train and validation datasets, while keeping authored eval cases in the official evals/evals.json layout. The runtime is single-shot: it returns one optimized prompt result and does not coordinate leader election internally.

When to use this skill

• The user says "optimize my prompt", "improve this prompt", or starts with /trainer-optimize
• The target artifact is a single .md (markdown) prompt template file
• The user provides, or can provide, train and validation JSONL datasets with expected outputs

Do not use this skill for general code optimization or non-markdown prompts.

Required inputs

The skill always needs a prompt path. It also needs train and validation datasets. In this repository, authored eval cases live under evals/evals.json, while APO datasets are passed explicitly.

• prompt: path to the markdown prompt file to optimize
• train: path to the JSONL training dataset
• val: path to the JSONL validation dataset