Autoresearch

Generate 10+ variants of a piece of conversion copy, score every variant with a five-judge simulated panel, evolve the top performers through a few rounds, then hand the winner to a human for the call — never past that point on its own. No traffic needed, no separate tooling: this runs the same way every other skill in this repo does, as instructions applied directly in this conversation.

What this replaces: staring at a blank headline field, or writing three tired variations by hand and picking the one that "feels right." What this doesn't replace: ab-testing — a strong simulated score gets you a better starting candidate to test, not a validated result. Run this first, then validate the winner with ab-testing against real traffic.

Before Starting

Check for existing strategy context first: If .agents/marketing-strategy.md exists, read it — Section 10 (Brand Voice) and Section 9 (Customer Language) matter directly here: a variant that scores well but doesn't sound like the brand is not a win. Also check .agents/marketing-learnings.md for past entries on this content type or channel.

Confirm before running:

1. The content — paste it, or point to the file/page.
2. Content type — product/landing page, email subject or body, ad copy, SMS, or signup/form page. Ask if it's ambiguous.
3. Which elements to optimize — headline only, or the full set for that content type (below)? Default: all.
4. Minimum score to stop at — default 80 (see Quality Gates).

The Loop