sammo

Why we included this project

SAMMO treats prompt writing as a search problem rather than a manual craft, which is useful once hand-tuning stops paying off. It represents a prompt as a structured program instead of a flat block of text, then applies mutation operators like paraphrasing or rewriting to generate candidate versions and scores each one against your own accuracy metric. A beam-search optimizer drives the search, and the library handles parallel execution and rate limiting, so you can sweep large batches of prompt variants against an LLM API without writing the orchestration glue yourself. You can start from a familiar Markdown prompt and use CSS-selector-style queries to target specific sections for modification, which keeps the workflow approachable. It's a good fit for developers and ML engineers doing reproducible, data-driven prompt optimization for classification and extraction work, including structured outputs, rather than a general-purpose agent or chat framework.