darwinian-evolver

To optimize a prompt using an evolutionary search loop, provide a starting prompt and an evaluator that scores outputs against a trainable set and a holdout set. The LLM-driven loop iteratively proposes and evaluates improved candidate artifacts.

Fitness function based LLM evaluation scores candidate artifacts against measurable signals like exact match, regex pass-rate, SQL correctness, unit tests, or LLM-judge outputs. This scoring guides the evolutionary search loop to iteratively improve the target artifact.

You detect overfitting when tuning SQL or regex by defining trainable and holdout failure cases within your evaluator. The evolutionary search loop checks improvements against both sets to ensure artifacts generalize instead of only matching exposed examples.

Yes, you need to install specific dependencies to run LLM-driven evolutionary optimization. The skill requires the openai, jinja2, and pydantic libraries, and operates by running the upstream darwinian_evolver via CLI or a subprocess driver.

Yes, you can use an evolutionary search to improve small code snippets and unit tests. The optimizer accepts small code as an LLM-facing artifact and uses your fitness function scoring, such as runtime metrics or pass/fail checks, to evaluate and mutate candidates.