LaunchDarkly Experiment Guide Skill
An Agent Skill that takes a developer from "I should probably A/B test this" to a concrete, ready-to-create LaunchDarkly experiment — even if they've never designed an experiment and don't have a metric, a flag, or a hypothesis yet.
Overview
This skill teaches agents how to:
- Scan the codebase and git history to infer what the product is and what the team has recently built (experiment candidates).
- Scan the connected LaunchDarkly account for existing flags and metrics, so it reuses instead of duplicating.
- Propose a hypothesis and a primary metric grounded in that context — with worked examples — instead of asking the user cold.
- Wire up the flag-evaluation and metric-tracking code the user needs.
- Draft the full experiment and create it only after explicit confirmation.
It is the novice-facing, guided front door to experimentation. For the
mechanical lifecycle (assemble/start/stop with a winner) it delegates to
launchdarkly-experiment-setup, and
it reuses the metric and flag skills for those pieces.
How it fits with the other skills
| Job | Skill |
|---|---|
| Guided, end-to-end experiment creation for a non-expert | this skill |
| Mechanical experiment assembly + iteration lifecycle | launchdarkly-experiment-setup |
| Choose / create a metric, instrument the event | launchdarkly-metric-choose · -create · -instrument |
| Create / wrap a flag to match the codebase | launchdarkly-flag-create |
Installation
From your project's root directory:
npx skills add launchdarkly/ai-tooling --skill launchdarkly-experiment-guide -yPrerequisites
- A local codebase (the skill reads files and
git log). - To create the experiment automatically, either the LaunchDarkly MCP server configured, or a LaunchDarkly API token for the REST fallback. Without either, the skill still produces the full draft and the exact commands to run.
Usage
The skill activates when you ask about experimenting or A/B testing, especially when you don't yet know what to test:
I want to A/B test something in my app but I'm not sure whatHelp me set up an experiment for the new checkout flowWhat should I experiment on, and how do I measure it?