Growth Engine applies Karpathy’s autoresearch pattern to marketing channels. Its experiment engine creates A/B or multivariate experiments with a hypothesis, variable, variants and primary metric, logs data points as content is published, and scores each experiment with a bootstrap confidence interval and a Mann-Whitney U test.
An experiment moves from running to trending to keep or discard. A winner needs p below 0.05 and at least a 15 percent lift, and it is promoted automatically to a living playbook that you check before creating new content. Batch mode handles up to ten variants, and the engine can suggest the next variable to test. Two more scripts produce a weekly scorecard and check campaign pacing against targets.
What You Get
- Tracked experiments with statuses running, trending, keep and discard
- Statistical scoring with bootstrap confidence intervals and Mann-Whitney U
- A living playbook of proven winners per agent
- Suggestions for the next experiments
- A weekly scorecard in Markdown across channels
- Pacing alerts for lead and candidate targets against pipeline, recruiting and email APIs
Common Use Cases
- Test subject lines or hooks across several variants and keep only significant winners
- Check the playbook for proven rules before drafting new content
- Review a weekly scorecard across all channels
- Watch daily lead and weekly candidate pacing and get an exit code when behind
What Makes It Unique
Winner decisions come from statistics with explicit thresholds you can tune, such as p-value, minimum lift and samples per variant, with fewer samples needed for high-volume agents like content and email. Winners flow into a playbook that is meant to be read before the next piece of content, so learning compounds.
Under The Hood
It is script-led with three Python scripts that need numpy and scipy. The experiment engine stores experiments and the playbook as files in a data folder you configure. The pacing script calls your own pipeline, recruiting and email platform APIs over HTTP using bearer tokens that you provide as environment variables. On start the instructions also call two telemetry scripts that log usage locally, with remote telemetry only by opt-in; those scripts are not in this skill folder.