Experiment skills - GrowthBook Documentation

These Agent Skills take an A/B test from idea to decision, all from your AI agent. Install the plugin and run /growthbook:gb-setup first. The example prompts can be typed verbatim, or your agent may run these skills automatically when it detects matching intent.

GrowthBook's Head of Experimentation, Luke Sonnet, runs an experiment end to end using the skills.

Prerequisites

A datasource and at least one metric, configured in the GrowthBook UI. The skills reference metrics by ID. They don’t create datasources or metrics.
The experiment skills target standard A/B tests. Multi-armed bandits aren’t supported yet: the skills stop and point you to the app rather than misinterpret a bandit.

The experiment lifecycle

The five skills below form a chain from ideation to decision. Each one routes to the next when you’re ready.

Skill	What it does	Example prompt
`experiment-brainstorm`	Propose new test ideas grounded in your team’s past stopped-experiment history. Read-only.	”What should we test next on the checkout page?”
`experiment-design`	Walk through hypothesis, variations, primary metric, guardrails, and sample size to produce a launchable spec. Read-only.	”Help me design an A/B test for the new pricing page”
`experiment-launch`	Create the experiment, create or reuse the feature flag, wire the experiment-ref rule, and start it.	”Launch the pricing experiment we just designed”
`experiment-analyze`	Refresh the snapshot if it’s stale, then interpret results: lifts, confidence intervals, guardrails, and a data-quality (SRM) check. Read-only.	”What are the results of `exp_abc123`?”
`experiment-stop`	Stop a running experiment, optionally declaring a winner and rolling it out to 100% of eligible traffic.	”Stop the pricing test and ship the winner”

Related concepts: Experiment configuration, Experiment results, Decision framework.

From experiment to cleanup

After you stop an experiment, the feature flag it ran through usually still needs tidying: inlining the winning value at your code sites and removing the flag. Hand off to flag-cleanup, which walks through code-site inlining and a two-step archive-then-delete.

Next steps

Feature flag skills: create, target, and clean up the flags your experiments run on
Agent Skills overview: install and setup
Experimentation documentation: the underlying concepts

Join our community Slack for additional tips and tricks