aiconfig-variations

安装量: 80
排名: #9815

安装

npx skills add https://github.com/launchdarkly/agent-skills --skill aiconfig-variations
AI Config Variations
You're using a skill that will guide you through testing and optimizing AI configurations through variations. Your job is to design experiments, create variations, and systematically find what works best.
Prerequisites
Existing AI Config (use
aiconfig-create
first)
LaunchDarkly API access token or MCP server
Clear hypothesis about what to test
Core Principles
Test One Thing at a Time
Change model OR prompt OR parameters, not all at once
Have a Hypothesis
Know what you're trying to improve
Measure Results
Use metrics to compare variations
Verify via API
The agent fetches the config to confirm variations exist API Key Detection Check environment variables — LAUNCHDARKLY_API_KEY , LAUNCHDARKLY_API_TOKEN , LD_API_KEY Check MCP config — If applicable Prompt user — Only if detection fails Workflow Step 1: Identify What to Optimize What's the problem? Cost, quality, speed, accuracy? How will you measure success? Step 2: Design the Experiment Goal What to Vary Reduce cost Cheaper model (e.g., gpt-4o-mini) Improve quality Better model or prompt Reduce latency Faster model, lower max_tokens Increase accuracy Different model (Claude vs GPT-4) Step 3: Create Variations Follow API Quick Start : POST /projects/{projectKey}/ai-configs/{configKey}/variations Include modelConfigKey (required for UI) Keep everything else constant except what you're testing Step 4: Set Up Targeting Use aiconfig-targeting skill to control distribution (e.g., 50/50 split for A/B test). Step 5: Verify Fetch config: GET /projects/ { projectKey } /ai-configs/ { configKey } Confirm variations exist with correct model and parameters Report results: ✓ Variations created ✓ Models and parameters correct ⚠️ Flag any issues modelConfigKey Required for models to show in UI. Format: {Provider}.{model-id} — e.g., OpenAI.gpt-4o , Anthropic.claude-sonnet-4-5 . What NOT to Do Don't test too many things at once Don't forget modelConfigKey Don't make decisions on small sample sizes
返回排行榜