darwinian-evolver

Define an Organism holding your prompt template, an Evaluator that scores outputs against expected results, and a Mutator that asks an LLM to propose improvements from failure cases. Then run the EvolveProblemLoop for several iterations and inspect the ranked snapshots.

Copy the shipped custom_problem_template.py and replace the artifact field with your regex_pattern or sql_query. Implement the run method to execute the pattern against test inputs, then score matches in the Evaluator and let the mutator propose fixes.

Yes, the skill ships scripts/parrot_openrouter.py which routes LLM calls through OpenRouter using the OpenAI SDK, so any OpenRouter model works. The upstream CLI itself is hardcoded to Anthropic and requires ANTHROPIC_API_KEY.

Provider content filters, especially Azure-backed OpenRouter models, reject phrases like ignore previous instructions. Wrap the LLM call in try/except and return an error marker string so the evolver scores that organism zero and continues.

Avoid it when the optimization target is differentiable, where gradient descent or DSPy fits better, or when you only need two or three variants that you can write by hand. It also fails when the fitness signal is purely subjective with no measurable criterion.

A typical run uses 50 to 500 LLM calls depending on iterations, parents per iteration, and concurrency settings. On gpt-4o-mini that costs pennies, while Claude Sonnet runs can cost a few dollars.