This is for running structured AI model experiments through LaunchDarkly's feature flag infrastructure. You clone a baseline config, override just the model or prompt or parameters you want to test, and compare variations systematically. The core workflow is disciplined: change one thing at a time, measure it, keep your baseline safe. It's useful when you need to reduce costs by testing cheaper models or improve quality without guessing, and you want the comparison data in one place. The clone tool inherits everything you don't explicitly override, which enforces good experiment hygiene. Requires their hosted MCP server, so this is really for teams already in the LaunchDarkly ecosystem who want to treat model selection like a feature rollout.
npx skills add https://github.com/launchdarkly/agent-skills --skill aiconfig-variations